Selected Engineering Work

Problem: Recruiters describe roles in free-form chat, and matching them against 160K+ candidate profiles had to be both relevant and fast.

• Built the agent that converts a recruiter's chat into a structured hiring spec over a 106-field job schema.
• Co-architected embedding recall followed by LLM reranking, with online preference learning from like/skip feedback and cold-start exploration.
• Migrated retrieval to pgvector and added caching plus an incrementally refreshed inverted index, cutting insights latency from 22s to 0.2s.

Problem: AI customer support for 3 enterprise clients across 60+ properties had to handle WhatsApp/WeCom traffic without losing messages or failing replies under load.

• Designed a modular agent workflow engine (Spring Boot + LangChain4j) with reusable nodes and conditional routing for intent recognition, parallel multi-query RAG over Milvus, and MCP tool calls.
• Built multi-tenant message ingestion with Redis per-tenant locks, TTL deduplication, and DB-backed cursor recovery.
• Added bounded thread pools, timeouts, retries, circuit breaking, fallbacks, and RabbitMQ buffering, raising AI reply success from 83% to 96% in 200+ concurrent load tests.
• Processed 11,800+ tickets and cut ticket creation time from 2 minutes to under 30 seconds.

Problem: An AI design-to-order product needed a backend that stayed fast under load and kept model costs under control.

• Built and deployed React + FastAPI microservices on AWS covering design generation, product previews, and ordering, validated with 500 simulated end-to-end workflows.
• Added Redis caching over PostgreSQL, reducing database reads by 69% and API latency by 32%.
• Moved checkout onto an Amazon SQS pipeline, cutting P95 checkout latency by 77% at 100 concurrent users with under 1% failures.
• Guarded AWS Bedrock calls with request validation and duplicate-request protection.

A modular Python framework for skill-based LLM agents with swappable skill acquisition, retrieval, execution, and governance components. Includes a reproducible ALFWorld evaluation pipeline that measures task success, token cost, and component compatibility.

A statically exported Next.js site served from a private S3 bucket through CloudFront (Origin Access Control, TLS via ACM) with Route 53 DNS. A CloudFront Function rewrites clean URLs to static pages.