// pipeline
EduAgent Pipeline
cache_router short-circuits on cache hit, skipping all LLM calls; rewrite generates N variants (2-5) by complexity, verify and evaluate run in parallel; AgentState as shared context, zero direct node coupling.
_
Chemistry major with CS/Math minor, pivoted to Java backend and AI systems engineering. Built production OCR-LLM-evaluation pipelines during internship — engineering AI into business systems: async task orchestration, cache service interfaces, workflow state machine control, production integration.
// about
// projects
Netty Boss-Worker Reactor API gateway. 6-stage filter chain (Auth/RateLimit/CB/GrayRelease/Log/Route), Disruptor Lock-Free Queuepeak-shaving, dual rate limiting (Guava + Redis Lua), Resilience4j CB + Nacos dynamic routing.
Community forum for 10K+ DAU.Caffeine + Redis Two-tier Cacheboosting post list QPS 20x, Kafka async reducing main API latency 40%, ES + IK tokenizer for ms-level full-text search.
Distributed LRU cache in Go, benchmarked Get 7ns/op zero allocation. Consistent hashing + virtual nodes + SingleFlight + Protobuf. 7-day progressive build.
LangGraph 11-node workflow:OCR → LLM Rewrite → SymPy Verify → Evaluate, async orchestration + cache APIs + state machine control, engineering AI into production.
Education content platform. 8-stage state machine for OCR-rewrite-review, 6-role RBAC per transition, Celery async for OCR + LLM tasks, AI integrated via service APIs.
Cross-process deterministic hashing, solving Python hash() inconsistency. Stack-iterative MD5, dict-order invariant, LRU cache, 1.8-2.5x faster.
LLM evaluation tool. PDF import → AI auto-grading → 16-sample averaging → variance analysis, helping partners optimize models with targeted data.
Desktop OCR tool. PyQt5 UI + QThread async, MinerU OCR extracts questions from images, LLM generates solutions and evaluations. Batch processing supported.
// pipeline
cache_router short-circuits on cache hit, skipping all LLM calls; rewrite generates N variants (2-5) by complexity, verify and evaluate run in parallel; AgentState as shared context, zero direct node coupling.
// skills
// experience
// contact