About

Owning backend architecture for a healthtech platform: the agentic booking system, the abuse defence in front of it, the worker infrastructure behind it, and the AWS bill under it.

• Built a patient-facing booking agent using a 17-tool calling loop matching doctors to complaints and handling payments, with typed error envelopes for in-loop model recovery.
• Layered abuse defence ahead of inference via pre-model extraction screening, an IP-banning strike tool, OTP quotas, HMAC-SHA256 session binding, and a 20-key PII redaction pipeline.
• Closed the loop on agent outcomes with a reaper that classifies expired conversations via a 4-field structured LLM call, with retry accounting and retroactive status correction.
• Built an AI session summariser generating doctor notes and patient-specific summaries; audio is transcribed via Deepgram ASR and indexed in Pinecone (Gemini) for RAG.
• Led extraction of the notification and worker service from the FastAPI monolith; a 4-queue Redis Queue system with per-channel selective retry (FCM, Exotel SMS, SMTP) decouples lifecycle events and reminders from the request cycle.
• Cut slot-query latency to sub-50ms across 200+ endpoints with a MongoDB-backed doctor-availability cache using compound indexes and explicit invalidation.
• Rightsized ECS-on-EC2 fleets (Auto Scaling, Aurora read replicas), cutting AWS spend 25% at a 99.9% uptime SLA.

Remote contract work, split between product delivery for SMB clients and evaluation work for AI labs.

• Shipped 8+ production full-stack apps (React, Node.js, PostgreSQL) for SMB clients; owned architecture, CI/CD, and deploy lifecycle.
• Reviewed 200+ LLM code solutions for Scale AI (Outlier); authored RLHF QA rubrics driving a 12% lift on coding benchmarks.
• Engineered adversarial Docker terminal benchmarks (AfterQuery) exposing multi-step reasoning gaps in SOTA models; built reproducible evaluation harnesses.