I build systems that hold under load.

I own the architecture end to end. Clinical consultations are captured as audio, transcribed, scored and reported back to ophthalmology practices, so every layer touches protected health information, and every failure has a patient behind it.

The ingestion path is built to survive its worst day: chunked 30s writes to IndexedDB, resumable upload with retry and exponential backoff, server-side FFmpeg reassembly, real-time WebSocket streaming. It survives mid-session client crash and network loss with no unrecoverable sessions.

Underneath: SQS with dead-letter queues, idempotent consumers and poison-message handling, holding zero sustained queue backlog. PHI is sealed with AES-256-GCM envelope encryption through AWS KMS, isolated decryption boundaries, and PostgreSQL row-level security enforcing multi-tenant isolation on every read path.

Concurrency is handled optimistically rather than with a distributed lock. A recovery sweep claims a consultation by writing a claim token, and every later write asserts both the processing status and that the token is still the one it read, because status alone proves a row is claimed and never that the claim is ours. Idempotency is enforced at the database: unique constraints, with an explicit lost-race branch on constraint violation rather than an error path.

Two schedulers write the same columns: the Lambdas, and six pg_cron jobs running inside Postgres. Report listings use keyset pagination over a non-unique sort key, so the cursor is a composite of date and id; a single-column cursor silently skips rows at page boundaries. The queue is tuned the same way, with a visibility timeout set to six times the function timeout and a dead-letter queue fanned in from 56 functions.

I also built a hardware recorder integration end to end: a native iOS client in Swift, WiFi SDK device pairing and file sync, and the path that moves on-device audio into the PHI-safe transcription flow.

Sole engineer. Generative product photography for e-commerce. Nine tools on OpenAI GPT-Image-1, deployed serverless with scale-to-zero and cold-start tuning under bursty load. An automated brand-extraction pipeline (scraping, image processing, GPT-4o-mini) ran with bounded concurrency and graceful degradation against unreliable third-party sites.

I owned billing correctness end to end: idempotent webhook processing, exactly-once plan provisioning, credit-based metering under concurrent writes. The failure mode there is charging someone twice, so it gets the same treatment as a claims path.

Led architecture and delivery while directing four engineers and one designer, owning technical decisions and release scope.

Built the PDF and SOP ingestion stack (chunking strategy, vector indexing, natural-language querying), cutting manual troubleshooting time 60%. Plus agentic workflows for troubleshooting, predictive maintenance and contract analysis, with retry, fallback and tool-failure handling across multi-step chains.

I rewrote legacy banking applications from scratch in Java 17 and Spring Boot 3.3, decomposed monoliths into microservices with real service boundaries and inter-service authentication, and integrated PingFederate for federated identity.

I led the Spring Batch migration off legacy VMs onto Red Hat OpenShift, and the wider platform move from Pivotal Cloud Foundry, covering NFR compliance, container orchestration, and the CI/CD cutover from UCD and Jenkins to GitHub Actions and Harness. Moving Java services from Ehcache to a centralised Redis cache, owning invalidation and connection pooling, cut response times 21%.

I fixed severe security vulnerabilities under Project Glasswing consistently ahead of remediation deadlines, and built the test estate from nothing with Playwright and Karate. A Maven-to-Gradle converter I wrote in Python was adopted across teams and removed 100+ hours of manual migration.