BharathStaff AI Engineer · Production Multi-Agent Platforms
← All case studies

Enterprise AI platform · 2025 –

Enterprise Modular RAG (Sole Owner)

Senior AI Engineer, Platform (sole owner, RAG service)

RAGGraphRAGMCPMulti-tenantCompliance

Challenge

Enterprise customers needed tenant-safe retrieval across unstructured documents: graph-aware search, governed tool access, and hardening for compliance products. A single shared index and ad-hoc prompts were not acceptable.

Approach

Tech stack

Modular RAG pipeline · GraphRAG · Vector DB · BM25 hybrid retrieval · MCP · Fastify · Prisma · Postgres · GCP Secret Manager · LangChain · Langfuse

Outcomes

Verifiable patterns (redacted)

Tenant collection model (redacted)

Each org gets isolated vector collections and graph partitions keyed by tenant ID. Ingestion pipeline validates org scope before write; retrieval MCP tools reject queries missing tenant context. Auth boundary: OIDC token → org claim → collection allowlist.

Hybrid retrieval config (redacted)

GraphRAG + BM25 + dense vector fusion with tenant-scoped reranker. Query path selects retrieval mode by intent classifier; all three indexes share org partition keys so cross-tenant leakage is structurally impossible.

Onboarding playbook excerpt (redacted)

Checklist for new org corpora: document classification, chunk policy, embedding model lock, eval golden set, MCP tool registration, and auth hardening sign-off. First customer onboarding completed in two sprints using this playbook.

Context

Compliance SaaS provider (customer anonymized): knowledge base for regulated product documentation.