Shubhang Subramanya

A RAG system built around one invariant: the vector index should never silently drift from the source of truth. Most RAG systems re-index on a batch schedule, so a document can change in Postgres and the answer engine keeps citing stale content indefinitely, with no signal that anything is wrong.

gRPC service with bidirectional streaming in async Python: a chat room with per-client queues and broadcast fan-out

A production RAG API that answers questions grounded in your own documents, built the way it would run in production: async FastAPI, PostgreSQL + pgvector (HNSW ANN), Redis cache, two-stage retrieval (vector search → cross-encoder rerank), containerised, deployed to Kubernetes via GitOps, and fully observable.

A transactional safety layer for multi-agent LLM systems - deterministic replay, cryptographic provenance, verifier - gated consensus, and prefix rollback.

A decentralized cloud cost optimization protocol that schedules container workloads across compute providers using blockchain escrow, priority-based scheduling, and real-time pricing oracles.

CertOps is an experimental autonomous DevOps system that simulates, verifies, and certifies infrastructure remediation actions before executing them in production.