Summary
Full-stack engineer taking AI products and developer tools from architecture to production. Co-owned a major product migration and delivered performance refactors at Merlin AI, and built automation, monitoring, and deployment tooling at Iterate AI. Work spans frontend interfaces, backend services, and infrastructure, with projects focused on resumable agent workflows, hybrid document retrieval, and isolated code execution.
Skills
- Languages: TypeScript, JavaScript
- Frontend: React, Next.js, TanStack Query, Zustand
- Backend: Node.js, Express.js, PostgreSQL, MongoDB, Redis, Docker, AWS
- ORMs: Prisma, Drizzle Vector DBs: pgvector, Pinecone, Qdrant
- AI/LLM: RAG pipelines, LangGraph, LangChain, Agentic tool orchestration
Experience
- Co-owned Bonkers end-to-end v2-to-v3 migration, unifying recreation, regeneration, and editing workflows into a single image‑to‑image pipeline, delivering daily stability improvements and weekly UX features, contributing to a 50% increase in DAU post‑migration.
- Developed Templates feature, enabling one-click generation of high-quality images across multiple styles and use cases, simplifying the creation process and resulting in 20K+ images generated in the first month.
- Executed a full-scale refactor of Merlin Chat, improving codebase readability, modularity, and maintainability through component abstraction and memoization, resolved critical memory leaks and boosted the Lighthouse score from 30 to 90+.
- Built automated code-generation tool for event tracking, analyzing developer workflows to create solutions that reduced manual implementation time by 70%.
- Contributed to deployment infrastructure overhaul designing scalable platform architecture that achieved 3x faster deployment cycles with reduced production incidents.
- Created internal monitoring dashboard for service health management, implementing real-time alerting system that cut incident resolution time by 50%.
Projects
- Built run admission for a prompt-to-production app builder: one Postgres transaction, serialized by advisory locks, checks global queue-depth, per-user, and per-chat caps before a run is enqueued to a BullMQ worker.
- Made agent runs resumable: typed stream events (files, commands, build status, preview URL) are appended to Postgres under a sequence number, so clients reconnect with Last-Event-ID and replay only missed events.
- Isolated generated code in per-chat Docker sandboxes that stay offline except during installs and builds, with Redis locks to stop concurrent requests from provisioning duplicates; previews ship to S3 and CloudFront via a Cloudflare Worker.
- Routed each chat turn to vision, document retrieval, mem0 memory, a pgvector semantic cache, or Composio tools (Gmail, Drive, Slack, GitHub), behind conversation-ownership checks, AES-256-GCM-encrypted user API keys, and a token budget.
- Built hybrid document retrieval in Postgres: tiktoken-sized chunks, HNSW pgvector and full-text search merged with reciprocal rank fusion, then Cohere reranking; retrieved text reaches the model as tagged untrusted data.
- Gated email and Slack sends and Drive, Calendar, and GitHub deletes behind LangGraph interrupt() approvals; Postgres checkpoints hold the paused thread, so a pending approval survives a page refresh and resumes in place.
Education
B.Tech in Information Technology (7.9 CGPA)