Ashwin Deepak Upadhyay

Architected a production-grade RAG engine with WAL-backed ingestion and strict namespace isolation, ensuring 100% data consistency across 500+ restart scenarios. Built sharded BM25 indexing with hybrid retrieval supporting 10K+ documents per namespace, achieving sub-300ms query latency through adaptive ranking and context windowing. Engineered robust observability with 2,500+ lines of automated tests, Prometheus metrics, and admin tooling for production deployment on Railway.

Led end-to-end engineering of a multi-tenant, matter-isolated RAG backend using FastAPI, Pinecone, PostgreSQL, Redis/RQ, and Nginx. Built hybrid retrieval with BM25 + Voyage embeddings + RRF fusion + cross-encoder reranking, improving accuracy from ~20% to 80–90%. Designed scalable ingestion for 100+ parallel documents with ZIP expansion (20+ formats), SHA-256 dedup, and metadata extraction. Reduced vector storage cost by 30–40% through document and chunk-level deduplication. Implemented folder-level BM25/vector isolation enabling clean retrieval for Claude MCP and N8N workflows. Delivered 2,000+ lines of automated tests covering ingestion, metadata, namespaces, folders, performance, and query correctness. Production deployment included RQ worker autoscaling, systemd automation, staging smoke tests, ingestion dashboards, and admin tooling.

Engineered a LangGraph-based multi-agent system with agents for data processing, risk scoring, anomaly detection, insight generation, collaboration, and visualization. Implemented RAG ingestion + LLM reasoning for natural language financial insights and report generation. Built fraud detection using Isolation Forest and supervised ML pipelines, with explainability and risk summaries. Developed FastAPI backend and Streamlit dashboard with real-time transaction streaming and alerting. Added caching, workspace collaboration, visualization APIs, and a unified test suite for reliability.