Work

Context

A multi-tenant compliance platform used by GRC managers and auditors, who feed its answers into audit documentation. A confident-wrong answer reaches a regulator.

Built

• Hybrid retrieval: Qdrant (ModernBERT + BM25, RRF fusion) plus Neo4j graph traversal across compliance hierarchies.
• LangGraph agentic chat with intent routing across six intents and CRAG hallucination grading.
• Multi-signal confidence tiering surfaced to the user, with refusal when the evidence is thin.
• Custom RAGAS eval harness with LLM-as-judge; Langfuse tracing end to end.

Result

Multi-tenant scoping resolved at query time, with no per-tenant data duplication. Every answer is graded and traced before it reaches a user.

Context

A property-management SaaS where the agent writes SQL against the customer's own database, so a wrong query becomes a wrong business decision.

Built

• LangGraph agent over a custom MCP server for read-only database access, with schema caching and FK detection.
• MongoDB Atlas vector search, Qwen3 embeddings, and a structured business-rule context library.
• Accuracy CI-gated against golden datasets, so a regression fails the build instead of reaching a user.
• FastAPI SSE streaming, Langfuse observability.

Result

The read-only boundary is enforced in the policy layer, not by trusting the model. Accuracy is CI-gated, so a regression fails the build before it reaches a customer.

Context

Document processing and decision support in front of customers filing real paperwork, across multiple data sources.

Built

• Multi-agent system for document processing and decision support across multiple data sources.
• RAG over business documentation, with GraphRAG/KAG experiments for retrieval precision.
• Multi-layer memory (short/long-term, procedural) for cross-session context.
• Led the monolith-to-services migration and set the development standards the team adopted.

Result

Cut manual data entry by 65%. Mentored 8 engineers on AI/ML practice and production deployment.