SUMMARY
Backend-focused software engineer with a GenAI/RAG specialization, working in TypeScript (Strict), Node.js, Express, Fastify, and PostgreSQL. Designed, built, and deployed 3 solo personal projects—an agentic contract analyser, a grounded RAG document service, and an autonomous support-agent backend—gated by 560+ automated tests (318 + 249 across two repos) in GitHub Actions CI.
TECHNICAL SKILLS
Languages & Backend: TypeScript (Strict), JavaScript (Node.js, Bun), Express 5, Fastify, Next.js, REST APIs, SSE Streaming
Databases & Caching: PostgreSQL, pgvector, Redis (Upstash REST & ioredis), BullMQ Job Queues, Prisma ORM, SQL
GenAI & Retrieval: Deterministic Quote Verification, Hybrid Search (BM25 + Dense Vectors), ReAct Agents, Tool Calling, Multi-Model Failover, Evals & Benchmarking, Langfuse Tracing
DevOps & Testing: 560+ Automated Tests (Unit, Integration) + Adversarial Golden-Set Evals, GitHub Actions CI/CD, Docker Compose, Sentry, Pino, Vercel, Render
AI-Assisted Dev: Claude Code, Cursor (Autonomous Workflows, Spec-to-Code Generation)
FEATURED PROJECTS
- Blocks displayed-citation hallucinations by construction with a deterministic quote-verification engine that resolves every model answer to source byte offsets and page coordinates before render — unverified text cannot reach the UI, and every displayed quote is click-to-highlight auditable.
- Accelerated contract research across up to 5 documents simultaneously by designing an autonomous ReAct research agent loop with a hard 4-round cap and server-side trusted tool execution.
- Automated financial risk detection in contract amendments, creating clause-level comparison logic that flags high-risk numeric shifts (e.g., liability cap changing $100,000 → $1,000,000 as CRITICAL) regardless of identical phrasing.
- Ship-gated by 318 automated tests (301 backend + 17 frontend) running on every push in GitHub Actions (CI green as of Oct 2026); async 202 Accepted ingestion with Redis-backed rate limiting absorbs burst traffic without dropping jobs.
- Refuses out-of-scope queries instead of hallucinating via strict distance-gated retrieval (≤ 0.5); 14/14 golden-set eval cases passed (Sep 3, 2026 committed run), including adversarial trap prompts that catch parametric-memory leakage — every eval run is committed as JSON in-repo, so results are dated and auditable rather than standing guarantees; 249 automated tests gate every merge.
- Offloaded 10–30s document parsing & embedding operations from the main HTTP thread using an asynchronous BullMQ and Redis background queue worker.
- Streams responses token-by-token over Server-Sent Events (SSE), paired with multi-provider waterfall failover (Groq → Gemini) so the service keeps working through provider outages and rate limits.
- Automated multi-step customer support workflows (order tracking, ticket escalation, refund drafts) using a custom ReAct agent loop built in pure TypeScript with zero heavy framework bloat.
- Runtime Zod schema validation on every LLM argument blocks malformed and hallucinated tool calls before any Prisma write lands — 0 invalid writes across the CI suite.
- Sustained agent availability through provider rate-limits and outages with automated model failover routing (Llama 70B → 8B) and structured Pino + Sentry error telemetry.
EDUCATION
Vadodara, Gujarat, India
CGPA: 8.13 (First Class with Distinction) · Relevant Coursework: Database Management Systems (DBMS), Operating Systems, Computer Networks, Data Structures & Algorithms, Software Engineering