AI & technology in ophthalmology
Most of my applied work does one thing: it uses machine learning to make a prediction a surgeon currently has to make from a proxy measurement. Four projects, all doing that.
| Project | What it is | What it does |
|---|---|---|
| ICL Fit | Deployed AI sizing agent, previously published as VAULT-OCT. Co-founded, and I lead the engineering | Identifies the best-fitting EVO ICL size for an individual eye from its own anterior-segment imaging, trained on real surgical outcomes, used before a case rather than sitting in a paper |
| VAULT & VAULT-OCT | The peer-reviewed models behind it, Journal of Cataract & Refractive Surgery | Predict what each lens size does in a specific eye from anterior-segment imaging, instead of inferring the sulcus from white-to-white corneal diameter |
| IOL Reference | Intraocular lens technology platform. Built and maintained by me | Lets cataract surgeons compare lens designs, optical principles, materials, power and toricity ranges, constants and specifications side by side |
| EyeLabs AI | The company behind the work | Builds AI tools for modern ophthalmology |
IOL Reference is structured data rather than a model, and that is deliberate. Choosing an implant is the least instrumented step in modern cataract surgery, and before you can reason about lens selection computationally you need the lens data in a form a machine can read.
What connects all of it: the parts of eye surgery machine learning has actually improved are the ones where an outcome depends on anatomy nobody can measure directly. ICL vault is one. Intraocular lens power is another, which is why several of the best modern IOL formulas are machine-learning models trained on tens of thousands of real outcomes. The parts it has not improved are the counselling decisions, which depend on what a person does for a living and what they want from their vision. Where that line sits today.
Frontier AI: reinforcement learning and evaluation
I build reinforcement-learning environments and evaluation harnesses that train and measure frontier models, spanning agentic coding agents, tool use, and reasoning. The environments pose real, verifiable tasks in sandboxed execution with automated reward and verification (RLHF and reinforcement learning from verifiable rewards), so models can be trained and benchmarked at scale against ground truth.
The tasks run deep into ML systems and GPU work: CUDA, Triton and Pallas kernels, distributed training (FSDP, ZeRO, NCCL), optimizer internals, and high-throughput inference (vLLM), alongside full-stack software engineering and infrastructure.
Medical-AI evaluation
As a physician, I focus on the part where my two fields meet: evaluating medical AI. I build rubric-based health benchmarks in which frontier models are graded against physician-written criteria for accuracy, completeness, safety, and communication, with model-based graders validated against physician judgment. It is the same problem as the clinic: measuring performance honestly against the standard of an expert.
Focus areas
Related: ICL sizing & vault research · cataract surgery & IOL selection · blog · CV
← Back to home