Joe JoeChen0430

Hi, I'm Joe (Yi-Jhao) Chen 👋

AI Engineer · Backend & Distributed Systems

Boston, MA · MS CS @ Northeastern University · Open to Fall 2027 co-op (Jul–Dec)

💡 What I focus on

I build LLM systems that stay reliable in production: fault injection and graceful degradation, retry classification, grounding checks, and the distributed-systems plumbing underneath.

🚀 Featured Projects

⚙️ Distributed Task Scheduler

A mini-Airflow: DAG scheduling across multiple workers

Problem: run dependent tasks across workers without double execution or lost work when a worker crashes

Implemented:

✅ Atomic SQL claim instead of check-then-act (verified with a 3-worker contention test)
✅ Exponential-backoff retries & per-task timeouts
✅ Failure propagation to dependent tasks
✅ Heartbeat / lease reaper recovers work from crashed workers
✅ ~700 tasks/s on wide fan-out; bottleneck traced to Postgres round-trips and the single dispatcher

asyncio · PostgreSQL · Redis · FastAPI · React

Tests · CI · Benchmark report

🧠 Tiered-Memory RAG Platform

RAG that degrades gracefully instead of failing

Problem: keep answering and keep the index fresh when models, caches, or retrieval misbehave

Implemented:

✅ Content-hash incremental ingestion + blue/green index swap: single-doc update 14 min → ~8 s, no empty-index window
✅ Circuit breakers, time budgets, 4-stage fallback chain (primary → small model → stale cache → retrieval-only): 99.4% success under 30% injected failure
✅ Confidence policy engine routes low-grounding answers to human review: override rate 31% → 7%

LlamaIndex · ChromaDB · Ollama · FastAPI · Docker

💼 Industry Highlights — Tricuss (Software Engineer)

Hybrid-search RAG (Qdrant + LangChain + Azure OpenAI); a heuristic short-circuit cut unnecessary LLM calls by ~70%
Invoice OCR structuring on Azure OpenAI vision: 9–10 validated fields per document, ~80% field-level accuracy on a 200-invoice eval set
ETL pipeline ingesting Redmine data into a data lake with Apache Spark + Apache Iceberg

🔧 Open Source

modelcontextprotocol/python-sdk #1103 — reproduction & root cause

Reproduced the StdioServerParameters stderr failure on v2 main and traced it to stdio_client's errlog default being bound to sys.stderr at import time; proposed resolving it at call time.

📫 Contact

Open to AI Engineer / Backend co-op roles for Jul–Dec 2027. Reach me at chen.yijh@northeastern.edu or on LinkedIn.

Navigation Menu