⌘resume

ai engineer with a full-stack background shipping production llm systems: rag pipelines with hybrid retrieval, llm evaluation harnesses, and streaming chat interfaces. built and deployed 2 live ai products (fastapi + next.js) on gcp cloud run / vercel / supabase, including a rag system scoring 0.94 faithfulness under a custom llm-as-judge eval harness.

• metquay inc · software engineer (calibration-management saas for labs)

  jun 2025 - present
  • building an offline-first pwa companion that keeps lab customers working across multiple offline devices, with git-style conflict resolution for concurrent edits.
  • designed the offline-online sync layer: indexeddb + service-worker sync queue with conflict detection and resolution flow for edits made across devices.
  • built ci/cd automation with github actions for pr auto-checks and plane api integration for pr tracking.
• zerone consulting · software development intern

  jan 2025 - may 2025
  • built a tuition-management app with jwt role-based access control across 3 user roles (students, teachers, admins).

• indian startup ecosystem rag (isra)live appdemosource code
  • built a production retrieval-augmented generation (rag) system over 111 indian startups with hand-rolled hybrid retrieval: pgvector cosine search + postgres full-text search fused via reciprocal rank fusion (rrf), reranked by a bge cross-encoder — no langchain, full control over ranking and citation grounding.
  • engineered an llm evaluation harness with hit@k, mrr, and custom llm-as-judge: hybrid+rerank lifted mrr 0.688 → 0.750 (+9%) over vector-only, 0.94 faithfulness on the golden set; built a retrieval-lab ui comparing vector, hybrid, and hybrid+rerank.
  • shipped a streaming sse chat endpoint (fastapi) with progressive sources, inline [source n] citations grounded by mapping them back to retrieved chunks, user feedback persisted to postgres, and langfuse tracing on every request.
  • deployed and monitored in production: fastapi + bge models (384-dim embeddings) on gcp cloud run.
  pythonfastapinext.jspostgresql + pgvectorsentence-transformerslangfusedockergcp cloud runsupabase
• readoraai — retrieval-augmented chat interface for pdflive appdemosource code
  • built a retrieval-augmented generation (rag) system for pdf question answering: parses, chunks (1,000 chars / 200 overlap), embeds, retrieves, and streams answers grounded in the retrieved context through a context-only system prompt.
  • engineered embeddings on gemini gemini-embedding-001 (768-d) with asymmetric task types for query-chunk alignment; retrieval on pinecone with per-file namespaces, top-k filtering, and score thresholding, answers streamed from gemini 2.5 flash.
  next.jstypescriptpineconegoogle geminipostgresqldrizzleneonclerkvercel blobtailwindcss
• multi-llm client — unified async python client for openai, anthropic & geminisource code
  • built a unified async llm client shipping four surfaces from one core: importable library, interactive repl, and fastapi service — normalizing messages, streaming, token usage, and error handling across openai, anthropic, and gemini apis.
  • hardened for production: 40 pytest tests across client, providers, api, and repl; tenacity exponential-backoff retries with a custom exception hierarchy; pydantic v2 under mypy strict; compare() races prompts across providers concurrently.
  pythonasynciopydantic v2fastapiclicktenacitypytest

• vellore institute of technology (vit)2021 - 2025 · vellore, india

  bachelor of technology in electronics and communication engineering (cgpa: 8.29/10)

• kestra · contributor

  open-source workflow orchestration — built and merged a 1,300+ loc todoist plugin for task automation.

• hacktoberfest 2025 supercontributorview badge

  6+ merged prs in global open-source repos in 30 days.

• aws certified cloud practitionerview credential

  cloud services, security, and architecture.