AI EngineeringFeaturedPersonal Project · May — Aug 2026
eSamvad
AI Document Intelligence Platform (RAG)
A multi-tenant document intelligence platform that turns 1,000+ page document sets into grounded answers. A hybrid retrieval pipeline fuses dense vector search with sparse full-text search, reranks the fused result with a cross-encoder, and streams the LLM answer back token by token over WebSockets.
- 105
- REST endpoints
- 24
- multi-tenant services
- 335
- automated tests
- 1,000+
- page documents
- Built a hybrid RAG pipeline fusing dense pgvector search with sparse Postgres full-text search via Reciprocal Rank Fusion, then Cohere cross-encoder reranking down to a top-5 context window — keeping 1,000+ page documents inside LLM context limits
- Engineered an async ingestion pipeline on Celery/Redis — PDF conversion, text extraction with Tesseract OCR fallback, chunking, embedding, and dual pgvector + OpenSearch indexing — with backoff retries and idempotent reprocessing
- Designed pluggable LLM, vector-store, and reranker abstractions so providers can be swapped at runtime without touching call sites
- Streamed token-by-token answers over WebSockets with Django Channels, keeping long-running generations responsive
- Delivered 105 REST endpoints across 24 multi-tenant services with per-tenant AI quotas and Stripe usage metering, covered by 335 automated tests
PythonDjango 5.2DRFLangChainLangGraphOpenAIpgvectorOpenSearchCohere RerankCeleryRedisAWS S3Docker