Harikiran Jangam

About

Backend engineer who built and evolved a PHI-compliant RAG system (FastAPI + LangChain + embeddings/FAISS) for internal document search and summarization, delivering <400ms p95 latency at ~2,500 daily requests and measurable impact (30% faster investigations, +17% retrieval relevance). Demonstrates strong security and rollout discipline (RBAC/RLS/JWT, redaction/audits, shadow mode, dual writes, canaries) and a focus on reducing hallucination risk via grounded guardrails and confidence-based fallbacks.

Experience

Education

Key Strengths

• Designed and shipped PHI-safe RAG backend for internal document search/summarization
• Achieved <400ms p95 latency at ~2,500 requests/day with predictable cost/latency tradeoffs (FAISS vs managed vector DB)
• Improved operational outcomes: 30% faster investigations and 17% better retrieval relevance
• End-to-end security implementation: RBAC, tenant scoping, PHI redaction, audits, and grounded-response guardrails
• Built reliable, frontend-friendly FastAPI APIs with strict Pydantic contracts, retries/timeouts, and health checks
• Led low-risk migration from monolith to FastAPI using shadow mode, dual writes, reconciliation, feature flags, canaries, and rollback
• Proactively identified retrieval failure modes (duplicate docs/entity confusion) and added deduping, stricter filters, and 'not enough signal' responses to reduce misleading outputs
• Built and deployed a production RAG search/summarization system for internal operations teams
• Reduced manual investigation time and improved response speed via AI-assisted document search
• Performance optimization: precomputed embeddings, fast vector search, and caching to reduce latency
• Improved answer quality through chunking strategy, retrieval tuning, and reranking
• Reliability-focused approach: defined success metrics, tested edge cases, added logging and guardrails
• Production orchestration experience using Airflow for automated retraining and Kubernetes for scalable inference
• Effective collaboration with non-technical stakeholders using demos and simplified workflows