M.S. Information Systems · Northeastern University · Graduating Dec 2026

I build reliable AI agent systems and backend infrastructure — from multi-agent runtimes, tool execution, and context engineering to evaluation, observability, state management, and production delivery.

Open to Agent Infrastructure / AI Backend / LLM Platform / Software Engineering roles · Available full-time January 2027

• Agent runtimes & orchestration — multi-agent workflows, tool-result loops, bounded execution, HITL, cancellation, replay, and failure semantics
• Evaluation & observability — frozen-input benchmarks, replay, LangSmith, RAGAS, OpenTelemetry, pass@k, fault injection, and cost/latency analysis
• Stateful AI backends — PostgreSQL, Redis/Lua, Kafka, async workflows, concurrency control, SSE/WebSocket, and cross-session memory
• Production delivery — Docker, Kubernetes/ACK, Helm, GitHub Actions, AWS EC2, rollout validation, and regression testing

Open-source AI education platform where I work on agent runtime, tool execution, grounding, evaluation, and reliability.

• Evolved the in-class agent runtime from a Pi-based POC into a bounded native execution layer, separating Director-led orchestration, child reasoning loops, and host execution with explicit budgets, cancellation/timeouts, terminal-state handling, and late-event isolation.
• Architected a closed-loop Tool Harness across 13 native tools and 4 execution paths, standardizing idempotency, replay, cancellation, exception propagation, and authoritative tool-result continuation.
• Designed authoritative state coordination around RuntimeStore + CAS optimistic concurrency + idempotent replay, with browser E2E coverage for duplicate delivery, delayed results, cancellation, and races.
• Built a 100-case frozen-input benchmark and replay harness informed by 141,102 production calls, compared 10+ model/provider configurations, and validated a selected route at 106-way concurrency for 10 minutes with 755/755 usable outputs.

Also contributed rendering/reliability work upstream to HyperFrames, including an opt-in static-frame reuse path and regression coverage for input identity.

Languages Python · TypeScript/JavaScript · Java 21 · Go · SQL

Agent & LLM Systems Pi Core · LangGraph/LangGraph4j · LangChain4j · ReAct · Planner–Executor · MCP · Tool Harnesses · HITL · Context Engineering · RAG · LangSmith

Backend & Data FastAPI · Node.js · Spring Boot · Reactor · PostgreSQL/pgvector · Redis/Lua · Kafka · MongoDB · MySQL · Elasticsearch · Pinecone · SSE/WebSocket

Frontend React/Next.js · Vue 3 · TypeScript

Cloud, Reliability & Evaluation Docker · Kubernetes/ACK · Helm · GitHub Actions · AWS EC2 · Linux · OpenTelemetry · Prometheus/Grafana · RAGAS · Playwright · Replay · Fault Injection

Stateful agent backend focused on memory lifecycle, concurrency, failure semantics, retrieval, and evaluation.

• Diagnosed a Redis lost-update race where asynchronous LLM compaction could overwrite concurrent writes; designed atomic Lua prefix replacement with deterministic concurrency and real-Redis integration tests.
• Redesigned tool outcomes around a backward-compatible SUCCESS / PARTIAL / FAILURE contract with retry-once and exception fail-safes, raising flaky-tool success 69.7% → 91.2% across 12 × 500 simulations.
• Built conflict-aware cross-session memory with supersession, context-scoped exceptions, recurrence accumulation, and source provenance; benchmarked against full-history, retrieval, LWW, and extract-only baselines.

Java 21 Spring Boot Redis/Lua Kafka Python LangGraph4j

Natural-language requirements → generated web applications, paired with a reproducible evaluation and regression platform.

• Coordinated 6 agent roles through typed RequirementSpec, TaskGraph, and ReviewReport contracts with bounded refinement and rollback.
• Bridged token streams into Reactor SSE on Java 21 virtual threads with concurrent request isolation and explicit cancellation/error semantics.
• Built a Python evaluation harness with structural checks, Playwright browser smoke tests, health gates, failure taxonomy, token attribution, and offline replay; preserved 90% structural pass@3 while cutting CodeGen-stage input tokens by ~18% in an isolated 10-case × 3-run A/B benchmark.

Java 21 Spring Boot Vue 3 TypeScript Python Playwright Redis

Autonomous AI agent system integrating RAG pipelines with tool-calling agents.

• Hybrid retrieval: vector search (PGVector HNSW) + BM25 via RRF fusion, cross-encoder reranking (ms-marco-MiniLM-L-6-v2) via Python FastAPI
• ReAct hierarchical agent: BaseAgent → ReActAgent → ToolCallAgent → LingManus with stuck-state detection and human-in-the-loop escalation
• 8 tool callbacks + custom MCP Server (Pexels image search, Stdio/SSE transport) + Amap location MCP integration
• LangSmith tracing via OpenTelemetry · SSE streaming · CI/CD with Docker Buildx ARM→amd64

Spring Boot Spring AI PGVector Python FastAPI Vue 3 AWS EC2 LangSmith

Full-stack clinic platform with a RAG-based AI health assistant grounded in an internal medical knowledge base.

• RAG pipeline: PDF ingestion → SHA-256 dedup → chunking → OpenAI embedding → PGVector (HNSW) → hybrid retrieval → RRF → SSE streaming
• RAGAS evaluation: Faithfulness 0.896 · Answer Relevancy 0.952 · Latency 310ms
• Patient booking system + admin document upload pipeline + knowledge base management

Spring Boot Vue 3 PostgreSQL PGVector OpenAI AWS EC2 Nginx

Distributed system for benchmarking LLM reliability, latency, and output quality.

• Kafka-based async pipeline decoupling model inference (2–37s) from API responses
• Redis Lua atomic caching: latency 1,962ms → 0ms on cache hits
• Sustained 4,800+ req/sec, P99 328ms · retry strategies + dead-letter queues
• Elasticsearch prompt search · Prometheus + Grafana observability

Spring Boot Apache Kafka Redis CockroachDB Prometheus Grafana React

Distributed gateway handling authentication, routing, and traffic control across 20+ backend services.

• Redis-based rate limiting and circuit breaker · AK/SK signature authentication · Dubbo RPC integration

Spring Boot Redis Dubbo

• LinkedIn: duan-ling-9970ba350
• Email: duan.lin@northeastern.edu