I'm a product manager who can't stop asking “how does this actually work?”

Problem

Retrieval-augmented generation (RAG) system to ingests support tickets' full context, retrieves relevant historical data, and produces a structured draft diagnosis for human review.

Build

Designed and built a multi-agent LLM pipeline to classify inbound email, auto-drafts contextual replies, and notifies the owner, with per-email tracing of latency, token count, and classification confidence Engineered for cost and reliability: cut cost by routing classification to locally-hosted model; added cost/rate circuit breaker, crash-safe idempotency, and auto-reply-loop prevention; unit and E2E testing included

Problem

RAG-based support triage tool that helps a tier-2 support engineer diagnose tickets across four shared-infrastructure applications

Build

Reduced time-to-resolution from 2 hours of investigation to minutes for reviewing troubleshooting steps, by designing and building an offline retrieval-augmented-generation (RAG) system that drafts structured diagnoses and resolutions for support tickets Created technical design document (problem statement, phased roadmap with exit criteria, success metrics, and risk register), then implemented hybrid semantic + keyword retrieval, confidence scoring, and a human-in-the-loop accept/reject/modify feedback loop Prioritized build-vs-buy and considered data-residency decisions (vanilla Python; 100% on-device inference), with Phoenix observability and SQLite audit trail for accuracy tracking