@JongHangAICHANNELVisit Channel

# DIGITAL TWIN of the HUMAN HEPATOBILIARY SYSTEM — Part 2: Hepatic Vasculature Network In memory of a friend, an exceptional teacher, and a legendary liver-transplant surgeon, Dr. Tan Kai Chah—KC Tan. In May 1995, I had the privilege of performing histopathological work for a landmark liver-transplant procedure led by Dr. KC Tan at Subang Jaya Medical Centre, Malaysia. One of his earliest lessons has stayed with me ever since: “Vasculature is the biggest and most unforgiving devil in liver transplant.” In liver surgery, the challenge is not simply to identify a lesion. It is to understand its relationship to the intricate network of vessels that supplies and drains the liver. A small difference in position can determine whether a vessel can be preserved, whether reconstruction may be needed, or whether a proposed resection carries unacceptable risk. That challenge inspired the development of this hepatic vasculature digital twin. Using CT-based image segmentation, the system creates an interactive three-dimensional view of the liver envelope, the segmented vascular network, and the tumor mask. In this example, the model identifies an estimated vessel volume of 43.5 millilitres, a tumor-mask volume of 48.4 millilitres, and 131 vessel components within the image-derived vascular mask. The key feature is tumor–vessel proximity analysis. Here, the nearest measured distance from the lesion to the vessel surface is 0.6 millimetres. The lesion is therefore classified as close—within five millimetres of a vessel surface. The system can display these proximity bands directly in 3D, helping users immediately see where surgical attention may be required. The accompanying imaging-only checklist records no direct tumor–vessel contact on the masks, no strong imaging signal of vessel encasement, and no multifocal lesions near vessels in this case. However, the close geometric margin is flagged for review. This distinction matters: a lesion may be near a vessel without invading it, but close proximity still warrants careful evaluation by the multidisciplinary team. For surgeons, this digital twin can support pre-operative discussion of whether a peripheral lesion may be approached with a limited non-anatomic wedge resection or whether a more formal hepatic resection should be considered. It can also help communicate lesion-to-surface and lesion-to-vessel relationships during board review, operative planning, and patient consent discussions. For patients, the same visual model can translate complex imaging into a clearer explanation of where the lesion is located, why vascular anatomy matters, and why a particular surgical strategy may be proposed. Importantly, this is an investigational educational and planning-support prototype. The vessel labels in this dataset are undifferentiated: they do not distinguish portal veins, hepatic arteries, or hepatic veins. The results are geometric mask metrics, not a formal Couinaud segment analysis, a named-vessel reconstruction, a future liver remnant assessment, or a clinical resectability decision. The purpose is to make vascular anatomy more visible, measurable, and discussable—while ensuring that final decisions remain grounded in complete imaging, clinical expertise, and surgical judgment.

# DIGITAL TWIN of Human HEPATOBILLIARY SYSTEM Part 1 - Patient-Specific Surgical Planning Patient-specific digital twins of the human hepatobiliary system create virtual models of the liver, bile ducts, gallbladder, and surrounding blood vessels. Built from CT and MRI imaging, they transform complex anatomy into interactive three-dimensional representations that offer more precise anatomical evaluation than conventional two-dimensional imaging alone. These models are valuable because hepatobiliary anatomy varies significantly between individuals. The hepatobiliary digital twins shown in this video are supported by AI agents designed for anatomical evaluation, surgical planning, and hepatic disease assessment. Their knowledge base is informed by authoritative clinical and surgical references, including Blumgart’s Surgery of the Liver, Biliary Tract and Pancreas, the NCCN Guidelines for Hepatocellular Carcinoma, and Couinaud’s Surgical Anatomy of the Liver Revisited. AI-powered operative-planning and anatomical-metrics agents evaluate liver lesions and generate key surgical information, including liver volume, lesion localization, future liver remnant—or FLR—assessment, resection considerations, and resectability checklists. These tools can support clinical decision-making and surgical planning alongside expert medical judgment. AI-generated graphics can make complex imaging findings easier to understand for both patients and physicians. For patients, simplified visual maps can explain a planned right-sided liver resection in plain language, show that multiple small lesions are localized rather than involving the entire liver, and clarify why nearby structures such as the bile ducts and pancreas may appear in the illustration. For surgeons and other physicians, these schematics can support discussions of peripheral lesion location, lesion-to-surface proximity, and possible approaches such as non-anatomic wedge resection versus formal hepatectomy. Educational and planning support only; these graphics do not replace formal clinical assessment, vascular analysis, FLR evaluation, or surgical judgment. Data credit: Antonelli, M., Reinke, A., Bakas, S., et al. “The Medical Segmentation Decathlon.” Nature Communications, 13, 4128 (2022).

PRESERVE AND REVITALIZE using AI This video shows how an AI solution helps users preserve privately owned books, articles, and other physical intellectual materials by converting them into accessible digital formats. Many valuable works in science, mathematics, and engineering are out of print and difficult to borrow or purchase, while personal copies may be aging, fragile, or beginning to deteriorate. With this AI-driven approach, owners can create lasting digital records of these scarce materials for personal reference and preservation. Diagrams and graphics can also be recreated and enhanced using vision-language models, subject to applicable copyright permissions. AI can re-examine diagrams, add clearer labels, and create updated versions that incorporate current findings and publications while preserving the original for reference. This approach can also help contemporary authors revisit published materials, suggest edits, redraw diagrams and charts, and update images in light of the latest research. Image artifacts introduced during scanning can be automatically identified and removed, while historically, culturally, religiously, and archaeologically significant materials are preserved in their original form and interpreted using purpose-built language models. Contemporary authors can use this approach to revitalize published works by incorporating clearer graphics and diagrams, as well as timely updates and corrections, to create a more engaging reading experience.

# AI for THE OPEN SEAS Ep 5 (Final) — Turning News into Operational Intelligence News feeds extend maritime intelligence with early signals that may emerge before formal advisories: port disruption, regional conflict, piracy, extreme weather impacts, industrial incidents, sanctions, and supply-chain constraints. AI can continuously turn these unstructured reports into location-aware, risk-rated operational context—while retaining source, time, confidence, and citations so that news informs, but does not replace, verified navigational and regulatory sources. This fifth and final episode of AI for THE OPEN SEAS demonstrates how AI, geospatial and optimization enable smarter, safer transportation for oil and gas. A coordinated team of AI agents transforms live news feeds into operational intelligence. The News Feed Agent reads, summarizes, rates, and classifies each item for operational risk. A Filtering Agent identifies reports relevant to maritime transportation, port operations, security, geopolitical developments, weather, marine piracy, and supply-chain disruption. The Coordinates Agent extracts referenced locations and assigns latitude and longitude coordinates, together with confidence and source provenance. The Vessels Agent retrieves each vessel’s current navigation position from onboard GPS or verified noon reports, then calculates the great-circle distance between vessels and relevant events using the Haversine formula. These results are fused with voyage plans, cargo condition, maintenance readiness, procurement constraints, weather forecasts, and regulatory requirements. The system can then prioritize events by proximity, severity, and relevance—alerting operators to risks that may require closer monitoring, route or speed adjustments, port-call changes, or human review. News intelligence is treated as an auditable decision-support signal: every recommendation remains linked to its originating report, publication time, location confidence, and risk rationale. This completes the five-episode journey from verified operational truth to explainable, proactive maritime decision-making.

# AI for THE OPEN SEAS Ep 4 — Extending Operational Truth with Naval Intelligence Every voyage begins with a verified position. From the latest noon report, we know exactly where the vessel is, how fast it is travelling, the prevailing weather, and the condition of its voyage. But knowing where a ship is only answers one question. Modern transportation of oil and gas demand another: what is happening around the ship, and what should we do next? This is where Naval Intelligence Extension transforms operational awareness into operational intelligence. The noon report becomes the anchor for a continuously evolving picture of the maritime domain. Satellite AIS traffic, naval advisories, weather systems, piracy alerts, search-and-rescue activities, and geopolitical security intelligence are fused into a single operational context. Every intelligence source carries its provenance—who issued it, when it was observed, its confidence, and, where applicable, the vessel or authority that generated it—creating recommendations that are transparent, traceable, and auditable. Maintenance records reveal critical equipment defects. Procurement systems identify urgently required spare parts and consumables. Cargo-specific telemetry monitors LNG boil-off gas, inert gas systems, tank conditions, and other segment-specific operational parameters. Together, these internal signals determine not only where the vessel can sail, but how safely it can operate under changing conditions. At the same time, every voyage remains governed by the Law of the Seas. Document Intelligence continuously searches IMO, MARPOL, ISM, MLC, ABS class rules, and company operating procedures. The relevant regulations are automatically surfaced with citations to the exact document, section, and page—providing explainable compliance exactly when decisions are made. All of these streams converge within the Voyage Risk Evaluation Engine. The video shows a single operational picture where operational truth, naval intelligence, maintenance readiness, procurement planning, and regulatory compliance come together to support every decision at sea. Rather than presenting isolated alerts, AI evaluates the complete operational picture. It considers vessel position, surrounding traffic, weather forecasts, maintenance limitations, procurement constraints, cargo condition, regulatory obligations, and maritime security intelligence simultaneously. The system then produces risk scores and operational recommendations—whether to maintain course, adjust speed, reroute around hazardous areas, delay a port call, or escalate the situation for human review.

# DIGITAL TWIN of a REFINERY - Crude Allocation, Process Optimization and Fuel Blending Petroleum refineries are among the most complex industrial systems in operation today. Their economic performance depends on the ability to transform diverse crude feedstocks into high-value fuels while operating within stringent process, quality, and capacity constraints. This digital twin models an integrated refinery that purchases multiple crude oils and processes them through distillation, reforming, catalytic cracking, and blending operations to maximize profitability. The optimization framework is adapted from the classical refinery planning problem presented by H. P. Williams and implemented using the Gurobi optimization platform. The refinery begins by separating crude oils into light naphtha, medium naphtha, heavy naphtha, light oil, heavy oil, and residuum fractions through atmospheric distillation. These intermediate streams may be directly blended into finished products or routed through secondary conversion processes. Reforming upgrades naphthas into high-octane gasoline components, while catalytic cracking converts heavier oils into cracked gasoline and cracked oil, increasing the overall value extracted from each barrel of crude. Final products include premium gasoline, regular gasoline, jet fuel, fuel oil, and lube oil. Product quality specifications impose operational constraints, including minimum octane requirements for gasoline grades, vapor pressure limitations for jet fuel, prescribed blending ratios for fuel oil, and production targets for lube oil. Refinery operations are additionally constrained by crude availability, distillation throughput, reforming capacity, and cracking capacity. Mathematical optimization determines the optimal allocation of crude feedstocks, intermediate processing pathways, and blending recipes that maximize total refinery profit while satisfying all mass balance, yield, process, and product quality requirements. Linear programming techniques enable simultaneous consideration of thousands of operational decisions and their economic implications. The digital twin provides refinery planners with a quantitative environment for evaluating production strategies, understanding process bottlenecks, assessing capacity expansion opportunities, and identifying the most profitable utilization of available feedstocks. Generative AI further enhances decision-making by interpreting optimization outcomes, explaining trade-offs between product slates and process utilization, identifying limiting constraints, estimating operational risks, and recommending economically advantageous operating conditions under changing market scenarios.

# DIGITAL TWIN of an AIR CARGO FLEET Part 2 - Network Intelligence and Critical Hub Analysis Air cargo systems are more than collections of airports connected by flight paths. They are complex transportation networks whose structure influences efficiency, resilience, capacity utilization, and vulnerability to disruption. Using the same Seabury dataset provided through the Faculty of Aerospace Engineering at Delft University of Technology, the digital twin transforms movements of UPS, FedEx, and DHL incorporate operational attributes such as Air Freight Tonnes and route distance, enabling the network to be studied from both capacity and geographic perspectives. Network science techniques reveal the hidden topology of each carrier. Degree Centrality identifies the most connected airports, while Closeness Centrality highlights nodes capable of reaching the rest of the network with minimal transit effort. Betweenness Centrality exposes strategic transfer hubs that facilitate cargo flows between otherwise disconnected regions, and Eigenvector Centrality measures the influence of airports connected to other highly connected airports. Additional metrics, including PageRank, Harmonic Centrality, and Hub–Authority scores, provide complementary insights into the hierarchy and operational significance of individual facilities. Comparative visualizations allow the transportation architectures of UPS, FedEx, and DHL to be examined side-by-side, revealing differences in network concentration, hub dependency, and connectivity patterns. Weighted analyses based on cargo volumes and distances further distinguish airports that are operationally critical from those that are merely highly connected. By integrating graph analytics into the digital twin, supply chain planners gain a quantitative understanding of network resilience, identify critical nodes susceptible to disruption, evaluate redundancy within the transportation system, and support strategic decisions related to hub expansion, route redesign, and long-term fleet planning. Generative AI enhances these capabilities by interpreting network metrics, identifying emerging risks, explaining operational trade-offs, and recommending actions to strengthen supply chain robustness under changing market conditions.

# DIGITAL TWIN of an AIR CARGO FLEET Part 1 - Routing and Aircraft Assignments Air cargo is a key facilitator of global trade and economic development. The resilience and sustainability of modern supply chains depend heavily on this mode of transport to move goods efficiently from producers to consumers worldwide. Beyond commerce, air freight plays a critical humanitarian role. Immunization programs save millions of lives each year, and their success relies significantly on the timely delivery of temperature-sensitive vaccines. Air cargo enables these pharmaceuticals to reach their destinations under tightly controlled conditions using advanced logistics technologies and handling procedures. Data used to construct this digital twin was obtained from the Faculty of Aerospace Engineering at Delft University of Technology in 2022. The dataset, provided by Seabury, contains the Air Cargo Transport Networks of UPS, FedEx, and DHL, comprising approximately six months of recorded departures that were scaled to estimate a full year of Air Freight Tonnes (AFT) with a combined legs of over 2300. In the video, the transportation networks of these carriers are displayed side-by-side for comparative analysis. Network science metrics—including Degree Centrality, Closeness Centrality, Betweenness Centrality, Eigenvector Centrality, PageRank, and Harmonic Centrality—are used to characterize and visualize the relative importance of airports within each network. Given origin and destination airports, graph algorithms identify all feasible routes and determine optimal paths based on criteria such as minimum hops or shortest distance. Using Seabury aircraft fleet data, an assignment optimizer further minimizes total flight operations and uncovered AFT by assigning integer flight frequencies to aircraft types and network legs while satisfying annual cargo demand. Generative AI in the digital twin augments Supply Chain Manager by providing actionable insights, highlighting critical decision points, evaluating fleet and capacity implications, estimating confidence levels and risks, and recommending optimal next actions.

# DIGITAL TWINS OF ANIMAL MODELS FOR BIOPHARMACEUTICALS – From Mouse to Patient, with a Pit Stop at Charles River Animal models, particularly mice and rats, remain indispensable in biologics development. They provide critical insights into a candidate therapy’s safety, efficacy, pharmacokinetics, pharmacodynamics, and toxicity profile before progression into human clinical trials. As living systems, they enable researchers to investigate complex biological responses, immune interactions, and disease mechanisms that cannot yet be fully replicated in vitro. Data generated from rodent studies are essential for regulatory submissions and contribute significantly to reducing the risk of clinical trial failure. This video demonstrates the development of Digital Twins for mice and rats, where biologics researchers specify experimental parameters such as species, cohort size, control groups, dosing schedules, dose levels, target affinity, tumor burden, study duration, and sampling cadence. These virtual representations are informed by animal model specifications commonly employed by Jackson Laboratory and Charles River Laboratories, drawing from my experience handling mice and rats for monoclonal antibody development at Inno Biologics, Malaysia, in 2008 within a cGMP environment. In practice, even genetically homogeneous cohorts of mice do not necessarily produce homogeneous outcomes. Consequently, causal inference frameworks such as the Neyman–Rubin potential outcomes model can be incorporated into Digital Twin simulations to quantify treatment effects, characterize inter-animal variability, and better predict translational outcomes from animal studies to human patients. The future of biologics development lies in the convergence of Digital Twins, AI, and experimental biology, transforming animal studies into predictive, data-driven engines for human therapeutics. It also redefine preclinical development by enhancing translational confidence, supporting regulatory decision-making, and reducing costly late-stage clinical failures.

# REPURPOSE DRUGS Using AI – Introduction to the jonghang-cchem-dgrp-256 Computational Chemistry Embedding Model In the late 1950s, thalidomide was marketed as a sedative and anti-nausea medication for pregnant women. It was later found to cause severe limb malformations in thousands of infants, leading to its worldwide withdrawal in 1961. In 1964, an unexpected discovery emerged when Israeli dermatologist Jacob Sheskin administered a forgotten bottle of thalidomide to a patient suffering from leprosy-associated skin lesions. The remarkable clinical improvement sparked decades of investigation into the drug’s biological activities. Researchers subsequently discovered that thalidomide possesses anti-angiogenic properties, enabling it to inhibit the formation of new blood vessels. Ironically, the same mechanism implicated in its devastating effects during embryonic development also makes it therapeutically valuable in diseases such as cancer, where angiogenesis supports tumour growth and metastasis. Today, anti-angiogenic therapies, including monoclonal antibodies and small molecules, constitute an important class of anticancer agents. Inspired by the possibility of uncovering new therapeutic applications for existing compounds, I developed two computational chemistry embedding models: 'jonghang-cchem-dgrp-256' for drug repurposing and 'jonghang-cchem-bsim-256' for biosimilar evaluation. Traditional embedding models learn semantic relationships by transforming natural language into numerical vector representations. In contrast, jonghang-cchem-dgrp-256 was trained on more than 400,000 pharmaceutical and bioactive compounds. Rather than encoding words and sentences, the model represents molecules through atomic-level features and computational chemistry descriptors derived from molecular structure, force-field calculations, molecular interactions, and molecular dynamics simulations. The resulting embeddings capture chemical and biological similarities between compounds, facilitating the discovery of therapeutic opportunities that may be overlooked by conventional approaches. To demonstrate its application, a chemical compound represented as a SMILES (Simplified Molecular Input Line Entry System) string was converted into a computational chemistry embedding and projected into the learned vector space. Similar compounds were retrieved based on embedding proximity and subsequently analysed to identify known biological activities, including protein targets, associated genes, disease indications, and mechanisms of action. These relationships were ranked according to embedding similarity scores. Generative AI was further incorporated to retrieve and summarise the latest findings from PubMed relevant to the identified diseases and therapeutic targets. By revealing previously unrecognised connections between existing drugs, molecular pathways, and disease mechanisms, this model provides a framework for hypothesis generation in drug repurposing, with the potential to reduce the time, cost, and risk associated with traditional target discovery and early-stage drug development.

AI for BIOPHARMACEUTICALS CELL LINE DEVELOPMENT Part 2 – Scientific Knowledge Orchestration and Patent Readiness Assessment Developing mammalian cell lines for monoclonal antibody manufacturing requires expertise across multiple disciplines. Molecular biology and genetic engineering are used to design high-expressing clones, cell biology ensures genetic stability, analytical chemistry verifies product quality, bioprocess engineering enables scalable manufacturing, and data science applies AI to identify the most promising production clones. Integrating these disciplines is essential for developing productive, robust, and regulatory-compliant manufacturing cell lines. This video demonstrates how AI retrieve the relevant scientific knowledge to support the interdisciplinary cell line development workflow while orchestrating knowledge generated throughout the development process to assess a cell line’s readiness for patent evaluation. Genetic stability is assessed by monitoring productivity retention across passage generations, as visualized in the Retention vs. Passage Generation chart within the Stability Studio. Multimodal AI explains the productivity trajectories of individual clones and interprets their long-term stability. To support clone selection, Generative AI analyzes experimental results, evaluates evidence, identifies potential benefits and risks, and provides an acceptance-favorability assessment to assist scientific decision-making. A novel production cell line may derive its commercial value from unique genetic modifications, proprietary engineering methods, manufacturing characteristics, or its application in producing a specific therapeutic protein. A successful production cell line is often regarded as a company’s “golden goose”—a proprietary biological factory capable of generating billions of dollars’ worth of therapeutic antibodies over its commercial lifetime. Generative AI synthesizes evidence accumulated throughout the cell line development program to evaluate the maturity of the supporting scientific evidence and provide insights into the cell line’s readiness for patent filing.

AI for BIOPHARMACEUTICALS CELL LINE DEVELOPMENT Part 1 – Monoclonality using Machine Learning In 6th century BCE Greece, a storyteller named Aesop told the tale of a farmer whose goose laid golden eggs. Driven by greed, the farmer slaughtered the goose and lost his source of wealth forever. In modern biopharmaceutical manufacturing, the goose that lays the golden eggs is the cell line. Monoclonal antibodies, hormones, growth factors, cytokines, interferons, and coagulation factors are examples of therapeutic proteins known as biologics. Many biologics are manufactured using mammalian cells grown in bioreactors, with Chinese Hamster Ovary (CHO) cells being the most widely used production host. To manufacture monoclonal antibodies, the gene encoding the therapeutic antibody is introduced into CHO cells, enabling them to produce antibodies that bind to specific disease-associated antigens. Cell line development is the process of engineering and selecting a single stable, high-producing CHO clone for manufacturing a biologic. Novel cell lines are essential to meet the growing global demand for biologics. Existing CHO hosts may be limited in productivity, genetic stability, or product quality, increasing manufacturing costs and development timelines. Engineering new, high-expressing, genetically stable cell lines enables higher product titers, consistent critical quality attributes, and support for increasingly complex biologics. It also strengthens intellectual property, reduces reliance on legacy production hosts, and accelerates delivery of affordable, life-saving therapies to patients. This video demonstrates how AI can accelerate the development of novel production cell lines. The first application of machine learning is clone ranking. Once early screening data are available—including titer, specific productivity (Qp), peak viable cell concentration, cell viability, and stability metrics—the platform predicts which clones should advance to the limited and costly downstream evaluation stages. Ranking identifies the highest-performing clones, but it does not reveal how they relate to one another. To uncover similarities and differences among candidate clones, the platform performs multivariate data analysis using Principal Component Analysis (PCA) on the standardized screening dataset. PCA visualizes clusters of clones with similar phenotypic characteristics and highlights potential outliers, providing additional insight for clone selection. The most expensive mistake in the development pipeline is advancing a clone into resource-intensive fed-batch evaluation only to discover that it performs poorly. To reduce this risk, the platform uses supervised machine learning models to predict key manufacturing outcomes—such as final titer and specific productivity (Qp)—from early screening measurements, enabling more informed and data-driven decisions before large-scale experiments are performed.

# PATIENT-SPECIFIC DIGITAL TWINS of the HUMAN HEART This digital twin of the human heart is built from an open research cohort of twenty-four four-chamber models published by Strocchi M, Augustin CM, Gsell MAF, KarabelasE, Neic A, Gillette K, et al. (2020) A publicly available virtual cohort of four-chamber heart meshes for cardiac electro-mechanics simulations. PLoS ONE 15(6): e0235145. https://doi.org/10.1371/journal.pone.0235145. The dataset was created so researchers can run large-scale computational cardiac studies without needing direct access to patient CT scans, the presence of senior Cardiologists or Cardiothoracic Surgeons. Each case comes from a heart-failure patient in a CRT upgrade cohort. The anatomy was segmented from cardiac CT and meshed in three dimensions. The models include ventricular fibre orientations and were checked against electro-mechanical simulations, so they are suitable for more than simple visualization. In the viewer, you start from baseline anatomy. Major structures can be turned on or off — left and right ventricular myocardium, atria, great vessels, vein and LAA rings, and valve planes — so you can focus on the regions that matter for your question. On top of that anatomy, scalar fields can be overlaid. Universal ventricular coordinates describe position from endocardium to epicardium, from apex to base, and around the ventricle. A separate mode highlights the CRT pacing site on the RV endocardium. Two-dimensional cut planes — axial, sagittal, and coronal — complement the three-dimensional view, alongside histograms and cohort-level clinical context drawn from the paper. Together, this turns a static mesh archive into an interactive window into patient-specific cardiac anatomy and coordinate systems. Heart digital twins turn patient CT into interactive, simulation-ready models with fibres and pacing sites. They support personalized CRT and EP planning by letting clinicians explore anatomy and stimulation scenarios virtually—reducing trial-and-error in invasive procedures and accelerating evidence for tailored cardiac therapies and surgical planning. Acknowledgement: My appreciation to the 2 retired medical professionals, a cardiac sonographer and a cardiothoracic surgeons for their invaluable input on how to make AI and digital twins useful, meaningful, pratical and easily accessible for clinical practice and patient care.

This video demonstrates a Digital Twin of a Cardiology Clinic in a large public tertiary hospital with an adjoining outpatient center. A family physician performs an electrocardiogram (ECG) on a patient and invokes a multimodal AI application to interpret the ECG. A generative AI assistant then drafts a referral letter, which is reviewed and verified by the physician before submission. At the cardiology care unit, the attending cardiologist reviews the referral and orders one or more diagnostic investigations, including: * Cardiac Magnetic Resonance Imaging (CMR) * Cardiopulmonary Exercise Test (CPX) * Resting Transthoracic Echocardiogram (TTE) * Exercise Stress Echocardiogram * Electrocardiogram (ECG) * Genetic Testing * Ambulatory Holter Monitoring Following completion of the investigations, a team of specialized AI agents—each trained to augment a specific subspecialty of cardiovascular care—analyzes the results and generates clinical narrations. These agents have been trained using authoritative cardiology references, including the Mayo Clinic Cardiology textbook, Harrison’s Principles of Internal Medicine, Hurst’s The Heart, Braunwald’s Heart Disease, the Cleveland Clinic Manual of Cardiovascular Medicine, Chou’s Electrocardiography in Clinical Practice, and Abbara’s CT & MR in Cardiology. A graphical representation of is generated using multimodal AI to provide an overall view of the patient. An orchestration AI agent then collects, reconciles, and synthesizes these outputs into structured clinical domains covering: * Triage * Diagnosis * Management * Potential Errors & Inconsistencies * Potential Omissions Role-specific conversational AI channels present the information appropriate for patients, primary care physicians, cardiologists, nurses, and other members of the care team through a chatbot. ACKNOWLEDGEMENT I first gratefully acknowledge O’Sullivan, J. O., et al. (2026), A large language model for complex cardiology care, Nature Medicine, 32:616–623, for sharing the data that enabled the development of this AI application. My thanks to a retired cardiac sonographer in Malaysia for his invaluable input, advice on the mechanism and challenges of cardiac care in the real world in the development of this digital twin.

## Biomolecular Kinetics of Cellular Therapy & Cytokine Release for Hematologic Malignancy A digital twin is a virtual representation of a physical system that mirrors its state, simulates its behavior, and predicts future outcomes. This video demonstrates a digital twin of the human vascular and immune systems used to perform scenario screening and predict the outcomes of cellular therapies for hematologic malignancies such as leukemia. Chimeric Antigen Receptor T-cell (CAR-T) therapy is a relatively recent treatment developed to combat leukemia and lymphoma. The therapy begins by collecting T cells, a type of white blood cell, from the patient. These cells are genetically engineered to recognize and bind to cancer cells. The modified T cells are then expanded in the laboratory and infused back into the patient, where they seek out and destroy malignant cells. The mathematical model draws on principles from physics and chemical engineering by mapping the behavior of a chemical reactor onto human physiology. An effective binding coefficient was first determined experimentally and then used as an input to the digital twin. This coefficient was incorporated into a biomolecular kinetics model using the following analogies: * The pumping action of the heart corresponds to the reactor impeller. * The vascular and lymphatic systems function as a well-mixed reactor. * Dose-dependent interactions between CAR-T cells and cancer cells define the reaction kinetics. * Changes in body temperature reflect the heat-transfer and thermodynamic effects associated with these reactions. * Blood pressure influences hemodynamics and fluid-mechanical properties such as shear stress. This shear stress affects the effective binding constant between CAR-T cells and cancer cells, with elevated blood pressure reducing binding efficacy. Using these relationships, the digital twin enables in silico scenario screening across different tumor burdens and Cytokine Release Syndrome (CRS) risk profiles. This allows researchers to evaluate treatment strategies virtually, helping prioritize a smaller number of more targeted and informative in vivo experiments.

# AI for the OPEN SEAS Ep 3 - Port Calling using AI Agents Driving to a mall for shopping typically involves finding an available parking space. Upon entering, drivers collect a parking ticket at the gate, navigate the parking lot, search for an empty spot, and park their vehicle. If the parking lot is full, the worst-case scenario is usually having to wait briefly for another shopper to leave and free up a space. Finding a parking space for a large cargo ship, on the other hand, is a massive undertaking. Navigation and port-entry preparation, documentation and regulatory clearance, cargo readiness and stability, machinery and engine readiness, safety and security compliance, environmental regulations, and coordination with port agents are just some of the many obligations involved. Every port call brings together dozens of parallel activities—berth bookings, crew-change rosters, open purchase orders, disbursement accounts, HSEQ inspections, outstanding PMS defects, and the latest noon-report position. Each phase, from 72 hours before arrival through anchorage, all-fast, departure, and port-state-control readiness, demands a different set of actions and checks. Today, much of this work remains scattered across spreadsheets, email threads, and disconnected Bassnet screens. Port Ops treats each port call as a live operational event, combining voyage and port-call information with crew, procurement, finance, safety, maintenance, and telemetry data into a unified operational view. No human operator can keep every dependency, obligation, and exception in mind every hour of a 24/7 operation. An AI agent can. By selecting a port call, operators can allow AI agents to analyze the unified context, identify risks and dependencies, and proactively recommend prioritized actions, alerts, checklists, and Bassnet updates. The video demonstrates how AI performs a final “last-mile” check, highlighting obligations that remain incomplete, overdue, or at risk of being overlooked.

# AI for THE OPEN SEAS Ep 2 — Getting the Right Parts to the Right Ship, at the Right Port, in Time Navies solved resupply with brute logistics. A battle group never sails alone — dedicated supply and replenishment ships steam alongside the warships, and through a network of cables and fuel lines strung between the two hulls, they transfer fuel, ammunition, food, and spares while both vessels are still underway. The fleet rarely has to stop or divert. Commercial cargo ships enjoy no such luxury. A lone bulker or tanker carries everything it will need for the voyage, and when something runs out or breaks, it must wait for the next port of call to take delivery. Every spare and store must therefore be anticipated, ordered, and routed to the right harbour before the ship arrives. A vessel is a self-sufficient factory at sea. When a fuel injector fails mid-ocean or provisions run low, the nearest store can be a thousand miles and several weeks away. Procurement officers juggle thousands of spares and stores across a fleet, each with its own reorder point, lead time, and criticality — while the ship keeps moving toward its next port. Order too late and the vessel waits; order at the wrong port and the parts never arrive. Disconnected requisitions, vendor catalogues, and voyage schedules make this guesswork. AI turns that guesswork into a plan. By reading live inventory against reorder thresholds, it surfaces exactly what must be replenished and ranks it by urgency and PMS linkage. It anchors every need to the vessel's *next port call*, checks whether supplier lead times can realistically meet the arrival date, and consolidates orders to the best delivery hub. It then matches each requirement to approved vendors by category, KPI, and proximity — drafting purchase orders with the right approval level automatically. Geospatial intelligence is what grounds these decisions in the real world. Google's Geocoding API resolves each port to a precise city and coordinates; the Embedded Maps API shows planners the delivery location and supplier-to-port routes at a glance; and the Places API scans for ship chandlers around the upcoming port. Together they answer the question behind marine procurement: 'can the right parts reach the right ship, at the right port, in time?' — turning reactive ordering into confident, location-aware planning.

# AI for THE OPEN SEAS Ep 1 — Turning Noon Reports and Maritime Law into Operational Intelligence Every day at noon, each vessel reports its position, speed, weather, and fuel burn. Across a fleet of LNG carriers, tankers, and bulkers, these noon reports are the heartbeat of operations — yet they arrive as thousands of inconsistent rows, where a single fat-fingered coordinate or impossible consumption figure can distort voyage performance and charter-party claims. Reading them one by one is impossible; trusting them blindly is dangerous. This video shows how AI brings order to this stream. It normalises every report into a canonical record, applies segment-specific extensions for each vessel type, and flags anomalies — speeds that defy distance, positions that jump oceans, consumption that breaks physics. Plotted on an embedded Google map, each ship's track becomes instantly verifiable, turning raw telemetry into trustworthy insight. But operations never run free of the 'Law of the Seas'. A ship is bound by ABS class rules and a dense web of IMO, MARPOL, ISM, and MLC regulations — tens of thousands of pages that no crew can recall under inspection pressure. Maritime Document Intelligence solves this. Regulatory and class PDFs are vectorised into a searchable knowledge base, letting officers ask plain-language questions and receive answers grounded in the actual text, with citations to the exact rule, section, and page. Multimodal Cross-Document Semantic Search: officers query not only with text but with images — a photo of an equipment nameplate, valve, or structural diagram is embedded and matched against the corpora, returning the rules, specifications, and figures it relates to. Cross-document search then connects related obligations across sources, so one query surfaces every relevant requirement at once — at the moment of a port-state-control inspection or a compliance decision. Together, Noon Reports and Document Intelligence give ship managers a single pane of glass: verified operational truth on one side, defensible regulatory grounding on the other.

Designing bioreactors is inherently complex due to nonlinear interactions among microbial metabolism, nutrient availability, oxygen transfer, temperature, genetics, and reactor conditions that collectively influence productivity, stability, scalability, and final product yield. This machine learning regression model is designed to predict and optimize microbial product yield in a bioreactor by analyzing a large combination of biological, chemical, and operational parameters. The system enables researchers and process engineers to evaluate how different fermentation conditions influence production performance and identify parameter combinations that maximize yield in g/L. The model accepts multidimensional inputs across several categories. Carbon source parameters include substrate identity, molecular weight, concentration, and elemental composition. Reactor and environmental conditions include reactor type, working volume, growth media, temperature, and oxygenation state. Microbial strain information captures genetic background, engineered modifications, directed evolution status, and pathway optimization strategies. Product-related features include molecular composition, precursor pathways, enzymatic complexity, ATP requirements, and cofactor costs such as NADH/NADPH usage. A parameter sweep capability allows users to systematically vary selected variables across large numeric ranges to generate multiple experimental scenarios automatically. These generated combinations can then be evaluated by the regression model to estimate expected product yield without performing costly and time-consuming wet-lab experiments for every condition. The platform functions as an AI-assisted bioprocess optimization engine that supports rapid design-space exploration, fermentation tuning, metabolic engineering analysis, and experimental prioritization. By learning from historical experimental datasets, the regression model can uncover nonlinear relationships and interactions between reactor conditions, microbial genetics, and metabolic pathways that would be difficult to detect manually. Such systems are particularly valuable in industrial biotechnology, synthetic biology, precision fermentation, biofuel development, and biomanufacturing, where optimizing yield, productivity, and resource efficiency is critical for commercial viability and process scalability.

Why AI Is Becoming Essential for Value-at-Risk (VaR) and Compliance in Capital Markets In 1994, Orange County, California declared bankruptcy after suffering catastrophic losses from leveraged interest-rate bets during the bond market crisis. The turmoil exposed major weaknesses in financial risk management and accelerated the adoption of Value-at-Risk (VaR), pioneered by JPMorgan Chase as a unified framework for measuring potential portfolio losses — a methodology that later became a global banking standard. Today, VaR remains one of the most critical concepts in fund management because it directly impacts capital preservation, leverage decisions, portfolio construction, regulatory compliance, and ultimately whether a fund can survive periods of market stress. Value-at-Risk mathematics is complex, combining stochastic modeling, probability distributions, correlations, volatility estimation, Monte Carlo simulation, and extreme tail-risk analysis. This video demonstrates how AI can automate and enhance daily VaR operations: generating morning risk briefings, running stress tests, maintaining their auditable trails, performing multi-factor cost-of-capital adjustments, designing custom scenarios and compare them against historical library scenarios, enforcing mandate-level VaR, drawdown, and sector limits, and validating model quality using Kupiec POF and Christoffersen independence backtests, and so forth. When anomalies, exceptions, or violations are detected, AI agents autonomously search regulatory documents, filings, policies, and handbooks to extract relevant insights, identify best practices, and help risk teams respond rapidly and consistently. AI agents also prepare source references — including documents, sections, page numbers, and paragraphs — to support further investigation.

Interpreting financial charts is inherently challenging because they are lagging historical records, littered with stochastic market "noise," algorithmic manipulation, and cognitive traps like confirmation bias that warp perception. Addressing these significant hurdles is critical because charts are the primary interface through which market participants determine real-time asset pricing and execute high-stakes investment decisions. Without the requisite deep knowledge and practical experience to filter these distortions, investors risk misinterpreting value and making serious errors in capital allocation based on flawed signals. Mastering chart interpretation is therefore essential to transform raw, chaotic price action into a disciplined framework for making rational investment choices and accurately assessing valuation risk. This solution demonstrates how AI agents can analyze intraday financial charts to help investors, traders, regulators, and risk managers better understand discrete hedging behavior and portfolio rebalancing dynamics. By interpreting price action, volatility patterns, and market movements in real time, AI systems can summarize market conditions, identify potential hedge adjustment points, and suggest follow-up analytical workflows. Interpreting intraday price movements and determining when to rebalance positions require deep quantitative expertise, fast decision-making, and continuous monitoring of market conditions. Small timing differences in hedge adjustments can materially impact portfolio risk, trading costs, and overall performance. This creates significant challenges because asset prices can change rapidly between rebalancing periods, exposing portfolios to volatility, execution risk, slippage, and sudden market shocks. Based on user-selected objectives, subsequent AI agents perform deeper quantitative analysis to evaluate risk exposure, hedging efficiency, rebalancing timing, and potential decision pathways tailored for different investment and risk-management personas-investors, fund managers, regulators & quantitative finance analysts.

UNLOCKING BURIED INSIGHTS IN SEC 10-K FILINGS United States Securities and Exchange Commission (SEC) 10-K filings are regarded as some of the most credible and defensible documents in the capital markets. They provide a wealth of information covering a company’s financial performance, business operations, risks, strategy, governance, and market outlook. Analysts study these filings to determine whether a company is fundamentally strong, fairly valued, and capable of generating sustainable future returns for investment decision-making. However, using 10-K filings comes with significant challenges. Large companies can produce filings of hundreds of pages long, written in dense legal and accounting language that requires specialized expertise to interpret correctly. Important information may be buried deep within narrative sections, while different companies often describe similar risks using inconsistent terminology. This video illustrates how AI can address these challenges to help investors make informed decisions more quickly and efficiently. To achieve this, 10-K documents were divided into chunks, converted into embeddings (numerical vector representations), and stored in a purpose-built vector database. Investors provide decision criteria such as industries, geopolitical risks, outlook, corporate governance, and contingent liabilities in the form of plain text, Word or PDF documents, and images. Under the hood, AI systems transform and process these inputs into technical representations and embeddings to retrieve the most relevant document chunks from the vector database. The retrieved contents are then interpreted and paraphrased, while associated metadata such as page numbers, sections, source 10-K filings, ticker symbols, and company names are extracted and linked together. Disclaimer: This is an AI experiment for and does not constitute investment advice. Please consult qualified financial advisors before making investment decisions. This experiment used only 1,000+ SEC 10-K filings.

WHY AI IS ESSENTIAL IN ADDRESSING CHALLENGES IN CAPITAL MARKETS Investors rely on news for signals, yet face overwhelming volumes of information, inconsistent quality, and noise that obscures true market relevance. News is fragmented, often biased, and highly time-sensitive—by the time it becomes widely known, it may already be priced in. Interpreting sentiment, context, and second-order effects requires expertise, while conflicting narratives add further uncertainty. Distinguishing material events from hype—and aligning insights to specific investment strategies—remains a persistent challenge. This video demonstrates how AI can address these issues by ingesting and analyzing news in near real time, classifying signals into bullish, bearish, and neutral categories. It also identifies relevant tickers, entities, and key actors mentioned across news sources. These tickers are then enriched with real-time market data and key statistics, enabling informed selection. Selected assets undergo further analysis using Efficient Frontier techniques, where AI agents perform data alignment, estimate returns and variance, construct the global minimum variance portfolio, and identify the tangency portfolio. The resulting Efficient Frontier visualization provides investors with a structured, data-driven basis for decision-making. Note: The Efficient Frontier is a way to choose investments that balance risk and return. It attempts to maximize return for a given level of risk, or minimize risk for a given level of return. Imagine each portfolio as a point on a graph: risk on one axis, return on the other. The Efficient Frontier is the curve showing the best possible combinations—those that give the highest return for each level of risk. Portfolios below the curve are inefficient (too much risk for too little return). Investors typically aim for points on the curve, depending on how much risk they are willing to take. Disclaimer: This is an AI experiment, not an investment advice. Please consult a qualified financial advisor before making any investment decisions.

Perhaps more than any other discipline, biological processes span an enormous range of scales—from DNA molecules to entire ecosystems. These systems are not linear; they are allosteric, dynamic, and highly complex. Advances in mathematics, molecular dynamics, genomics, transcriptomics, proteomics, biophysics, statistical thermodynamics, bioelectronics, reaction kinetics, and simulation techniques have enabled the design and biomanufacturing of novel molecules, hormones, monoclonal antibodies, and enzymes for industrial and medical applications. For decades, progress in enzyme engineering and biomanufacturing has relied on iterative experimentation—trial-and-error cycles in the laboratory that can take weeks or months. While effective, this approach is resource-intensive and difficult to scale. This video describes how AI can be used to design novel proteins, including enzymes, by combining models such as AlphaFold 3 and ESM Metagenomic Atlas. These models can be used both collaboratively and comparatively to evaluate candidate designs for new enzymes and small molecules. Starting from a desired function, candidate protein structures can be generated using diffusion-based methods such as RFdiffusion, followed by sequence design using ProteinMPNN. The resulting candidates are evaluated for structural stability, catalytic constraints, and folding accuracy. Predictions are cross-validated across models and further refined through simulation and iterative feedback. Instead of testing thousands of variants experimentally, large numbers of candidates can first be evaluated computationally.

INTERPRET 12-LEAD ECG USING MULTIMODAL AI Why ECGs matter An electrocardiogram (ECG) records the heart’s electrical activity. It is one of the most important tools for detecting heart attacks, abnormal rhythms, and other cardiac conditions. Because cardiovascular disease remains the leading cause of death worldwide, accurate ECG interpretation is critical. Why they are hard to read Interpreting ECG waveforms accurately takes years of training. Hospitals generate large volumes of ECGs every day, and subtle but clinically important features are easy to miss—especially under time pressure or fatigue. In ambulances or rural clinics settings, specialist interpretation may not be readily available. How ECG AI works In this video, ECG images and their corresponding interpretations are converted into numerical embeddings and stored in a vector database. When a new ECG of unknown condition is analyzed, it is transformed into the same representation, and similarity search is used to retrieve the most relevant historical cases, ranked by relevance. Any two ECGs can then be selected for side-by-side comparison to examine their similarities and differences. This approach extends expert-level support to settings where cardiologists are not readily available—AUGMENTING, NOT REPLACING, CLINICAL JUDGEMENT.

AI AGENTS for ENZYME DESIGN In 2007, I had the privilege of building a biopharmaceutical manufacturing process for a biotechnology company in Nilai, Malaysia, focused on producing humanized murine monoclonal antibodies for colon cancer. Monoclonal antibody (mAb) production remains one of the most technically demanding workflows in modern biotechnology and immuno-oncology. It sits at the intersection of immunology, protein engineering, structural biology, computational biology, and bioprocess engineering—each a discipline in its own right. Moreover, effectively combating cancer requires integrating expertise across all major “omics” layers—genomics, transcriptomics, epigenomics, proteomics, metabolomics, and phenomics. Cancer is not driven by a single defect, but by a dynamic network of molecular and cellular changes. At the DNA level, genomics identifies mutations and structural variations that initiate disease. However, genes alone do not tell the full story. Transcriptomics reveals which genes are actively expressed, while proteomics shows how these signals translate into functional proteins—the true executors of cellular behavior. Beyond this, metabolomics captures the metabolic rewiring of cancer cells in energy utilization, and epigenomics explains how gene activity is regulated without changes to the DNA sequence. Each of these omics domains is a deep and highly specialized field, often requiring years of PhD-level training and extensive academic and industrial collaboration. Historically, efforts to integrate these layers to fully characterize cancer have achieved only limited success. Such an undertaking typically demands the scale, talent density, and interdisciplinary coordination found only in elite research institutions—on the order of multiple “Whitehead Institutes.” In this showcase, three specialized AI agent teams are orchestrated to replicate and augment this multidisciplinary capability: - Literature Intelligence Team – Continuously ingests and synthesizes scientific literature to support problem formulation, hypothesis generation, and identification of redundancy or novelty. - Enzyme & Protein Design Team – Leverages platforms such as RCSB PDB, UniProt, AlphaFold, BRENDA, ExPASy, SABIO-RK, RFdiffusion, ProteinMPNN, and ESMFold to design and optimize candidate biomolecules for downstream validation and bioprocess scale-up. - In Silico Simulation Team – Integrates molecular biology, physiology, biochemical and computational engineering models to simulate reactions and system behavior within a human biological context. - Financial Engineering Team (ideation stage) - discover new financing models for rare diseases This AI-native approach enables small, focused biotechnology teams to achieve outcomes that previously required the scale and resources of large research institutions. Most importantly, it creates a compelling financial pathway that draws scientific talent back into the lab to tackle rare diseases such as Gaucher Disease and Duchenne Muscular Dystrophy. Ultimately, this approach aims to make Kendall Square–level innovation ecosystems and MD Anderson–calibre institutions reproducible anywhere in the world.

INDISPENSABLE AI FOR CORPORATE FINANCE: REDEFINING ACCOUNT RECONCILIATION Account reconciliation—the process of matching internal financial records with bank statements, invoices, and ledgers—is one of the most critical yet persistently inefficient functions in corporate finance. It sits at the foundation of financial integrity. And yet, it remains one of the most manual, error-prone, and resource-intensive processes in the enterprise. Reconciliation at scale involves thousands—often millions—of transactions: payments, receipts, fees, and FX adjustments. While most are routine, every single entry must be validated with near-perfect accuracy. There is no margin for error. Data inconsistency is systemic. ERP systems, banks, and payment gateways generate fragmented, unstandardized data—frequently with missing references or ambiguous descriptions. Traditional automation reaches a ceiling with an outcome of 70–90% match rates are but the remaining represents the most complex and highest-risk cases: * Partial and split payments * Aggregated or batched transactions * Chargebacks and reversals * Human and system-generated errors Exception handling requires deep investigation across multiple systems, documents, and context layers is slow, mentally taxing and highly error-sensitive. The consequences are real resulting in misstated revenue, undetected fraud exposure, audit findings and regulatory penalties During month-end, quarter-end, and year-end closes, this pressure compounds—driving long hours, operational fatigue, and increased risk of error precisely when accuracy matters most. WHY AI IS NO LONGER OPTIONAL This is not a workflow problem. It is a cognitive scale problem. Human teams cannot efficiently process the volume, variability, and ambiguity inherent in modern financial data. Rule-based automation alone cannot resolve context-heavy exceptions. This video demonstrates two coordinated groups of AI agents purpose-built to address reconciliation at scale: DOCUMENT-LEVEL RECONCILIATION Beyond transaction matching, financial operations depend on validating relationships across documents—each requiring contextual understanding: * Quotation vs Purchase Order (PO) compliance * PO vs delivery and goods return notes * Three-way matching (Invoice vs PO vs Delivery Order / GRN) * Credit and debit note validation * Payment-to-document reconciliation AI agents interpret, correlate, and validate these documents with context awareness—eliminating manual cross-referencing. BANK ACCOUNT RECONCILIATION A dedicated set of AI agents continuously matches bank statement entries against general ledger cash accounts to identify and resolve: * Items in the banks not in ledgers; items in ledgers not in the banks * Outstanding cheques * Deposits in transit * Bank fees and charges * Errors and unexplained discrepancies

AI as YOUR TYRION LANNISTER and HARVEY SPECTER in HIGH-STAKES CONTRACT NEGOTIATION The strategic acquisition of silver for chemical manufacturing of Ethylene Oxide Silver has long been a symbol of achievement and elegance—glinting from Olympic podiums as medals and adorning wrists and necks as jewelry. Yet beyond its shimmer lies a quieter industrial role: deep inside chemical plants, silver works tirelessly as a catalyst in the production of essential compounds such as Ethylene Oxide, a building block for plastics, antifreeze, and detergents. For companies that depend on it, acquiring silver demands extraordinary capability in navigating a maze of legal, environmental, financial risks, labor issues, currency fluctuations, and international trade relations, resource nationalism, environmental, sactions. At the negotiating table, multiple suppliers are often involved. Negotiators must balance competing price structures, ensure consistent purity and delivery schedules, coordinate logistics across borders, and manage the risk that a delay from any single supplier could disrupt the entire manufacturing chain. Even the financial decision of whether to purchase silver outright or lease it through metal-financing arrangements can significantly affect a company’s profitability, making the stakes of negotiation even higher. In this video, I revisit a past industrial consulting engagement by building several teams of AI agents to provide an upper hand at the negotiating table. These agents carefully examine leasing contracts across multiple dimensions, including financial obligations, insurance and risk allocation, regulatory and compliance requirements, and technical datasheets. At the same time, leasing contracts from various suppliers are placed side by side for detailed comparison. The agents then generate recommendations to support negotiation strategy. The outcomes are ultimately organized into an Advantage Leverage Machine—a framework for scoring, prioritizing, and identifying negotiation leverage points.

AI-DRIVEN SPACE TECHNOLOGY SPINOFF FOR TERRESTRIAL INDUSTRIAL APPLICATIONS For many years, engineers at NASA grappled with the challenge of protecting astronauts from the crushing G-forces of launch and high-speed flight. In 1961, at NASA Ames Research Center, aeronautical engineer Charles A. Yost began experimenting with a strange cushioning material that behaved unlike any foam before it. It softened with body heat, slowly molding to the human form before gently returning to its original shape when pressure was removed. Originally designed to absorb shock and distribute weight in aircraft seats, this material eventually found its way into homes decades later—transforming into what we now know as memory foam mattresses and pillows. A small but remarkable everyday comfort born from the quest to reach space. Critics often argue that spending billions on space programs is a luxury when those resources could be directed toward more immediate social needs. This view overlooks the long-term economic multiplier effects of frontier research and the unexpected innovations that emerge as technological spinoffs. Yet creating successful spinoffs from space technology requires rare interdisciplinary talent and skills spanning engineering, artificial intelligence, materials science, product design, and commercialization—expertise that is rarely found within a single team. This week’s AI experiment explores a new approach: assembling an interdisciplinary innovation team in the form of AI agents. In the video, an engineer first identifies a piece of space technology with potential relevance to an industrial problem on Earth. A team of AI agents then investigates how this technology might fill existing capability gaps. Finally, another AI team carries out the design process—introducing improvements and translating the concept into a new terrestrial product.

I first gratefully acknowledge O’Sullivan, J. O., et al. (2026), A large language model for complex cardiology care, Nature Medicine, 32:616–623, for sharing the data that enabled the development of this AI application. The original groundbreaking work was carried out at Stanford Medical Center Cardiovascular Institute.. This weekend’s AI experiment addresses a critical challenge in healthcare delivery: the scarcity of subspecialist medical expertise. This issue is particularly acute in cardiology, where timely and accurate clinical management directly determines patient outcomes. This application leverages the Google Gemini 3 model to augment clinical decision-making in complex cardiology care. A twist is introduced by separating the AI-outcome into 5-domain clinical reasoning framework: Triage, Diagnosis, Management, Errors & Inconsistencies and Omissions. In addition, multimodal AI (Nano Banana) is employed to generate clinically meaningful infographics that translate complex data, reasoning pathways, and management considerations into intuitive visual representations, enabling clearer communication among clinicians, across multidisciplinary teams and patients. A conversational chatbot interface is incorporated to enable interactive, “ask-anything” clinical queries and exploratory discussions. This human–AI interaction paradigm supports clinical brainstorming, hypothesis generation, and, etc.

MEDICAL AI — AI for Multiplex Immunofluorescence in Histopathology Thirty-one years ago, I had the privilege of working as an intern under the mentorship of one of the most distinguished pathologists, Dr. Methil Kannan Kutty—a legend in the league of Stanley Robbins, Ramzi Cotran, and Vinay Kumar. About ten years ago, histopathology made a leap with a new technique called multiplex immunofluorescence (mIF) akin to moving from a black-and-white photograph of a single person to a high-definition image of a crowded party. In the past, pathologists could focus on only one biological marker at a time. The arrival of multiplex immunofluorescence fundamentally changed this. Multiple markers could now be visualized simultaneously within the same tissue section, with all cells preserved in their original spatial context. This made it possible to identify different cell types and biomarkers, observe how they interact, and understand their collective roles—especially in complex environments such as tumors. However, this power comes at a steep cost. Multiplex immunofluorescence requires highly specialized and expensive microscopes, imaging systems, reagents, and tightly controlled laboratory environments. The process is slow and labor-intensive, resulting in low throughput and limited scalability. Fast forward to 2026. The advent of generative and multimodal AI is once again transforming histopathology. This video demonstrates a purpose-built large language model analyzing standard Hematoxylin-Eosin–stained images to generate microscopic examination findings and diagnostic insights. A vision-language model is then used to detect and infer multiple biomarkers directly from these images—without relying solely on complex laboratory workflows.

MEDICAL AI — AI for Multiplex Immunofluorescence in Histopathology Thirty-one years ago, I had the privilege of working as an intern under the mentorship of one of the most distinguished pathologists, Dr. Methil Kannan Kutty—a legend in the league of Stanley Robbins, Ramzi Cotran, and Vinay Kumar. About ten years ago, histopathology made a leap with a new technique called multiplex immunofluorescence (mIF) akin to moving from a black-and-white photograph of a single person to a high-definition image of a crowded party. In the past, pathologists could focus on only one biological marker at a time. The arrival of multiplex immunofluorescence fundamentally changed this. Multiple markers could now be visualized simultaneously within the same tissue section, with all cells preserved in their original spatial context. This made it possible to identify different cell types and biomarkers, observe how they interact, and understand their collective roles—especially in complex environments such as tumors. However, this power comes at a steep cost. Multiplex immunofluorescence requires highly specialized and expensive microscopes, imaging systems, reagents, and tightly controlled laboratory environments. The process is slow and labor-intensive, resulting in low throughput and limited scalability. Fast forward to 2026. The advent of generative and multimodal AI is once again transforming histopathology. This video demonstrates a purpose-built large language model analyzing standard Hematoxylin-Eosin–stained images to generate microscopic examination findings and diagnostic insights. A vision-language model is then used to detect and infer multiple biomarkers directly from these images—without relying solely on complex laboratory workflows.

An AI platform for event planning — built on intelligence, driven by results. Meet your elite AI event team: • An Ideation Specialist to spark fresh concepts • A Venue Selection Expert to find the perfect location • A Conference Scheduler to design seamless agendas • A Marketing Strategist to amplify your reach • A Content Creator to craft compelling stories • A Social Media Manager to create viral moments • A Logistics Coordinator to bring people and equipment together flawlessly Powered by local intelligence. Work in your native language — and get results in your native language.

This is a problem every engineering team knows too well: review a new revision of a schematic before it is released for fabrication. The updated drawing looks almost identical to the previous version, and after a long manual comparison, the only change noticed is the fuse rating. Several subtle but critical updates are missed. Weeks later, during commissioning, motors fail to start, safety logic behaves unexpectedly, and the PLC no longer matches the wiring. The investigation traces the problems back to those undetected schematic changes. --- This team of AI Agents compares 2 revisions of the same drawings by removing much of the manual effort and risk involved in comparing engineering drawings across revisions. By automatically determining whether two drawings originate from the same source and identifying changes such as modifications, deletions, additions, or movements, it enables engineers to instantly understand what has evolved between versions. This improves design accuracy, speeds up reviews, reduces oversight of subtle changes, and strengthens traceability across the engineering lifecycle. Teams can focus more on engineering decisions and less on time-consuming visual inspections. For businesses, this means faster design cycles, fewer costly mistakes, and smoother handoffs between engineering, manufacturing, quality, and suppliers. Teams catch issues earlier, reduce rework, and move to approval and production with more confidence. The result is shorter timelines, lower risk, and a scalable foundation for AI-driven engineering operations.

Our recent breakthrough applies Multimodal AI to electronics engineering design through a coordinated system of specialized AI agents. First, a Vision Agent examines and analyzes electronic schematics to identify, classify, and interpret components and their functions. Next, a Knowledge Agent selects the most relevant references from a provided library, reads them, and retrieves the appropriate equations and design rules for the identified components. A Network Agent then maps all nodes and connections, constructs a circuit network diagram, identifies relationships between components, and labels them for downstream reasoning. The Solver Agent takes over by identifying the problems to be solved, applying the retrieved equations, and devising a step-by-step solution plan. It also generates structured tables explaining what is connected, how it is connected, and why. Finally, the Reporting Agent consolidates outputs from all agents into a comprehensive engineering report, automatically delivered as a Microsoft Word document. This pipeline significantly shortens design cycle time and accelerates time-to-market—a critical advantage both in competitive industries and in real-world defense applications.

This experiment pushes AI agents not just to their limits — but beyond them — to support one of the most complex and highest-stakes gambits in the industry: OIL-FIELD VALUATION - the most capital-intensive rolls of dice in the global economy. In 2025, the top 10 oil companies on the Fortune 500 alone generated a staggering $3 trillion in revenue, underscoring the persistent and seemingly insatiable global demand for fossil fuels. We are often reminded of the massive revenues and profits earned by large oil companies, but far less attention is paid to the extraordinary costs, technical complexity, and risks behind those numbers. Nowhere is the risk of failure greater than in the upstream exploration and production sector. The journey of oil production begins with geological and geophysical surveys to image subsurface structures that may potentially harbor hydrocarbon reservoirs. Exploratory wells are then drilled and evaluated using advanced wireline logging tools — some employing radioactive sources — to confirm the presence and characteristics of a reservoir. This is the turning point of decision making to decide whether to proceed with extraction, defer it, or abandon the prospect altogether. In this experiment, a multidisciplinary team of AI agents — acting as petroleum and reservoir engineers, production engineers, economists, news and political analysts, project and financial risk specialists, industrial engineers, legal and regulatory experts, senior management, and an investment committee — is assembled to evaluate two oil-field valuation scenarios: * Real options analysis — incorporating financial, capacity, and geopolitical parameters * Optimization-driven development planning, delivering results through production and extraction optimization, hedge strategy optimization, capital structure optimization, stochastic programming, real options optimization, and a supervising agent that selects and prioritizes the outcomes A specialized agent uses the problem statement as context to select, gather, analyze, and repurpose news for outputs as decision intelligence, providing important context that continuously feeds back into the decision cycle - an approach to use context to create context. Finally, an utility agent that works tirelessly to feed the decision making cycle in real time with crude oil prices.

Twenty-five years ago, I earned my Autodesk AutoCAD and 3D Studio MAX certifications from Universiti Teknologi Malaysia, at a time when engineering was rapidly shifting from manual drafting to digital design. CAD revolutionized the field by replacing pen, pencil, ink, and stacks of paper with precise digital models. Engineers could now create, edit, store, and reuse designs—making changes instantly instead of redrawing entire sheets—and drastically reducing paper consumption. Advanced simulation tools such as ANSYS further elevated the role of engineering expertise. These tools made it possible to perform structural, thermal, and fluid-flow analyses before any physical prototype was built, allowing engineers to predict real-world behavior, identify weaknesses, and optimize designs through multiple virtual iterations. While this greatly reduced development time and costly trial-and-error, effective use of these tools still depended heavily on highly trained engineers—professionals who understood design principles, materials, physics, and real-world constraints. Today, AI agents represent the next major shift. They can increasingly assist—or in some contexts, temporarily substitute for—the absence of experts. These systems can automatically explore design alternatives, detect anomalies in simulation results, suggest optimizations, and flag potential safety or performance issues. In environments where highly trained engineers are scarce, AI agents can provide a form of embedded expertise, guiding users through complex design problems, accelerating early-stage engineering work, and lowering the barrier to sophisticated analysis. In this illustration, a team of AI agents augments the expertise of mechanical engineers, cost engineers, and quantity surveyors by analyzing technical drawings to generate bills of materials, identifying suitable vendors, and sourcing prices from parts databases. They work together to perform input validation, fluid property identification, piping and valve sizing, hydraulic calculations, special-condition analysis, code and compliance checks, documentation and reporting, as well as optimization and recommendations. While they do not replace true engineering judgment, AI agents have the potential to extend expert capability, democratize advanced design tools, and enable innovation even when human expertise is limited or unavailable. AI is ready—big time—to take on engineering’s toughest challenges: design, economics, risk, optimization, and code and compliance.

AI-Driven CAR-T Treatment Strategies for Hematologic Neoplasms Simulation An experiment to push the envelope of AI: to explore whether an interdisciplinary AI Agents can work as a coordinated team to manage a highly complex environment in a clinical setting. Chimeric Antigen Receptor T-cell (CAR-T) therapy is a relatively recent treatment developed to fight certain blood cancers, such as leukemia and lymphoma. CAR-T therapy begins by collecting white blood cells called T cells from the patient. These cells are then genetically engineered to recognize and bind to cancer cells. The modified cells are expanded into the millions in the laboratory and infused back into the patient, where they seek out and destroy cancer cells with near-sniper precision. Metaphorically, you enter enemy territory, recruit ordinary citizens, train them into elite assassins, and send them back to eliminate enemy soldiers — a feat of medical ingenuity straight out of a spy thriller, evoking Carrie Mathison, Jack Bauer, and even a hint of Walter White. However, CAR-T therapy can trigger the release of cytokines in the patient’s body. Excessive cytokine release can be life-threatening, while too few CAR-T cells may allow the cancer to continue proliferating. Determining the correct dose is therefore a delicate balancing act. In this experiment, AI agents are tasked with administering and controlling CAR-T infusion to maximize cancer cell elimination while minimizing cytokine release. A mathematical model of a fed-batch chemical reactor is mapped onto patient physiology. The pumping of the heart represents the impeller; the vascular and lymphatic systems act as a well-mixed reactor; dose-dependent binding between CAR-T and cancer cells represents reaction kinetics; and changes in body temperature reflect the heat transfer and thermodynamics of these reactions. Blood pressure determines hemodynamics, which in turn govern fluid-mechanical properties such as shear stress. This shear stress influences the effective binding constant between CAR-T cells and cancer cells, with higher blood pressure reducing binding efficacy. Disclaimers: This is a hypothetical model to test the limits of AI Agents without considering evidence based findings. Demo of binding constants is based on previously published data for Gleevec-BCR-ABL binding. In conclusions, while AI agents — and AI as a whole—are not yet ready for this feat, they have demonstrated enormous potential.

Multiple AI Agents working together to work out design equations to tackle the design problem for lift generation from a cambered airfoil at zero angle of attack. Pre-analysis, problem statement verification, reference book identification, analytical solutions, numerical solutions, diagramming and python code generation tasks were delegated to various agents to solve the complex problems. These agents were taught to solve the problems and derive new equations iteratively and systematically. When they are stuck, a text book was provided for them to refer to to get the relevant equations and information to continue. They were also trained to pick the right book for the right problem. Lift coefficient is one of the most crucial parameters to be addressed to make high performance and safe aircrafts.

Multiple AI Agents working together to work out design equations to tackle the design problem for lift generation from a cambered airfoil at zero angle of attack. Pre-analysis, problem statement verification, reference book identification, analytical solutions, numerical solutions, diagramming and python code generation tasks were delegated to various agents to solve the complex problems. These agents were taught to solve the problems and derive new equations iteratively and systematically. When they are stuck, a text book was provided for them to refer to to get the relevant equations and information to continue. They were also trained to pick the right book for the right problem. Lift coefficient is one of the most crucial parameters to be addressed to make high performance and safe aircrafts.