MedTech

I architected and shipped Ocula AI at Equinox Ventures: a smartphone application measuring the pupillary light reflex, owned end to end from applied research through to a device in clinical use.

I trained and cross-trained YOLOv5 and YOLOv11 object detection models in PyTorch, benchmarking across architecture generations on AWS SageMaker and EC2 GPU instances to select the production model, then optimised inference to over 30 frames per second on consumer smartphone hardware — resolution and confidence-threshold trade-offs, post-processing, on-device deployment, and profiling to find where the frame budget actually went.

The harder half was the measurement layer on top of frame-wise detection: converting per-frame output into a calibrated pupil-diameter trace over time, from video captured in uncontrolled real-world lighting rather than a lab. I characterised failure modes against a reference instrument rather than reporting detection accuracy alone.

The result was validated against the clinical-standard NeurOptics PLR-3000 pupillometer and published: Khuu SK, He R, O'Brien BJ. Establishing the operating conditions of 'Ocula AI' in capturing the pupillary light reflex. Clinical and Experimental Optometry 109(4):679–688, 2026. PubMed

Sole engineer on ReportHub, a clinical reporting SaaS platform in production with two pathology laboratory customers. I inherited a rapidly-built codebase and brought it to production standard while shipping the features required to launch a third customer.

I own the structured-extraction pipeline: Claude via Amazon Bedrock extracting structured data from pathology reports that arrive as PDFs, with output validation and prompt evaluation held to production standard and measured against reference reports.

I built the initial proof-of-concept with the founding team — Terraform infrastructure for the API, the original Electron client, and Deepgram speech-engine integration — then the production evaluation system: defining metrics, assembling reference sets, and monitoring transcription performance in production.