At MIT I built graph and generative models that design macromolecules and peptides, then closed the loop with automated synthesis. At Google and Amgen I took the same ideas to genomics and drug manufacturing. Today I keep building in the open on GitHub.
Brings together what newborns see, hear and feel over the first two years of life, and asks whether small neural networks tuned to a given age can turn those measurements into a quick, inspectable simulation of sensory perception. Part of the Baby Senses Lab.
Digital twins usually start from CAD drawings and laser scans, which puts them out of reach for most non-profits. Is a public video tour enough to start one?
Frames of a public news tour of a food bank become Gaussian-splat captures through a world model, a scripted Blender model fills in the parts the video never showed, and OpenUSD ties both to operations data served from a serverless AWS backend. The twin runs in the browser and in NVIDIA Omniverse on EC2 GPUs. Built in seven days for about $1.54 of compute, with a reference architecture mapped to AWS, Azure and Google Cloud and real sensors priced at about $12 a month to run.
Factory planning ideas like MRP, routings and OEE are hard to show with plant data nobody can share. A famous kitchen makes them visible to anyone.
A 3D twin of The French Laundry's three-star kitchen running a full tasting-menu service. Covers are the build plan, 53 recipes are bills of materials and routings, and a discrete-event simulation plays out tickets, stations and bottlenecks minute by minute. Parallel coding agents built the product, data, simulation and 3D layers.
AI world models train on adult-quality data from day one. Infants start with blurry, muffled input and still learn fast. What does that staging buy a model?
One world model per sense (vision, hearing, touch, body position), each fed through an age-tuned filter and grown as it trains, then fused. The public site lets parents see and hear like a baby of any age; researchers get the findings, papers and code.
A family's lab results sit in PDFs and paper scans from different clinics, and seeing a trend usually means handing health data to someone's cloud.
An Android app that reads reports with on-device OCR, parses biomarkers with rules backed by a small local language model, and has a person confirm every value before it is saved. The release build has no internet permission at all.
Prospective PhD students choose advisors from rankings and cold emails rather than research fit.
Drop in a CV or paper and a small language model running in the browser on WebGPU extracts research interests; embeddings then rank faculty across the top 100 universities. Nothing leaves the device. A scheduled pipeline is rebuilding the faculty index from open scholarly data.
“Digital twin” means something different in every meeting. A shared, cited history gives teams a common vocabulary.
An open dataset of milestones from andon lights and Apollo's simulators to AI world models and agentic twins, with a walkable 3D factory shown across six eras and research trends from OpenAlex. MIT licensed for courses and talks.
Most AI-and-jobs estimates count tech jobs. The larger exposure sits in everyday tasks across the whole economy.
Adapts the Iceberg Index method to India: generative-AI exposure scores mapped to national occupation codes and weighted by 2023–24 labor survey data, district by district, with an interactive dashboard and a technical paper.