AI Engineering Consultant · Aspect
Prototype
?
Simulated end-to-end trace with mock engineers — the production Navigator Copilot ran LLM intent detection, the 3-pass fuzzy resolver (93% routing accuracy on a 29-question benchmark), and a min-max-normalized weighted ranking engine with automated red-flag detection (⚠ = first-time fix below 80%).
Highlights
- Multi-Agent Orchestration: Architected and built a production-grade multi-agent LLM orchestration microservice using LangGraph, LangChain, FastAPI, and Groq (Llama-3.3-70b), enabling natural language querying of engineer KPI data across 18 performance metrics and 6 scoring categories.
- Ranking Engine: Implemented a weighted composite ranking engine featuring min-max normalization, configurable penalty thresholds, and automated red-flag detection, serving role-based leaderboards for 3 access tiers (engineers, trade managers, and senior leadership).
- Routing & Entity Extraction: Designed an intelligent routing layer with LLM-powered intent detection and a 3-pass fuzzy name resolution algorithm (exact → substring → sequence matching), achieving 93% intent-routing accuracy on a 29-question QA benchmark run and reliable entity extraction from ambiguous natural language queries.
- Async Pipelines & Caching: Built an async data pipeline with namespaced L2 caching (4 TTL tiers), semaphore-gated concurrency control, and structured JSON recovery from LLM outputs, drastically reducing redundant backend calls and optimizing stability under concurrent loads.
- Visualization Pipeline: Designed and deployed an end-to-end visualization pipeline bridging the full LLM agent → FastAPI → React chart flow, complete with cloud database persistence and conversation-reload rehydration across the backend, service layer, and frontend.
- Access Control Architecture: Introduced a dual-audience agent design separating self-scoped and administrative data flows, and enforced route-level access guards backed by automated test coverage.
- Diagnostic & Engineering Operations: Root-caused legacy chart-pipeline failures, API name-resolution bugs, and query timeouts arising from connection-pool saturation under concurrent async fetch load.
- Frontend Refactoring: Unblocked production builds by rewriting ~50 broken imports following a massive directory restructure, and implemented a transitive import-graph regression test suite to prevent future structural breakages.
- Testing & Quality Assurance: Delivered a 37-test suite with a 100% pass rate on a live end-to-end workflow, closing out the contract with 210+ automated backend/frontend tests passing cleanly.
Technologies
- LangGraph
- LangChain
- FastAPI
- Groq (Llama-3.3-70b)
- React