Contract engagement with Netflix's Security, Privacy Assurance & Corporate Engineering team. Built internal web apps for 1,000+ employees on the team's React stack, extended internal APIs with OpenAPI specs, and automated deployment via Spinnaker and Jenkins. The defining constraint: write code as if you won't be there to fix it — because you won't. That permanently changed how I treat documentation and API design as upfront design constraints rather than afterthoughts.

A production deep-learning fraud pipeline that taught me the gap between a model that performs and a model that ships. Ensembled image (EfficientNet) and text (BERT) classifiers into a gradient-boosted risk scorer, served sub-80ms via FastAPI on AWS ECS with CloudWatch drift monitoring. The asymmetry between missed detections and false alarms drove every threshold decision — high-confidence scores auto-actioned, mid-range routed to human review.

Fully serverless AWS architecture: Lambda functions per domain, API Gateway routing, DynamoDB with carefully modelled access patterns, S3 for company profile and document assets. React/Redux/Tailwind/SCSS frontend connecting to a stateless API layer. Three user roles — founders, judges, org admins — with distinct permission models and UI flows.

Deep multi-role React architecture. Startup profiles, judge scorecards, admin competition management and round progression — all shipped as a cohesive interface built on React, Redux, Tailwind CSS and SCSS with a fully serverless AWS backend.

The challenge was false-positive tolerance — in healthcare, the cost asymmetry is brutal. I'd spend more time on uncertainty quantification next time. The serverless deployment architecture was genuinely clean though.

The most intellectually honest work I've done. Self-supervised learning on medical data forces you to confront how much we still assume in standard supervised pipelines. Uncertainty quantification isn't a nice-to-have — it's the whole point in clinical contexts.

NLQ accuracy is deceptively hard to measure. We had solid precision metrics but the qualitative gaps only showed up in user testing. I'd instrument intent-recognition failures earlier next time — quantitative metrics alone are insufficient.

The no-code-to-code migration was more than a technical lift — it was a trust problem. I had to prove the coded version would be faster to iterate on. The bundle size drop was the key metric that won the argument.