Computer Science undergraduate graduating in 2027, focused on reliable AI systems, agents, retrieval, model routing, robotics, and developer tools.

I like turning probabilistic model behavior into systems that are testable, observable, and explicit about their limits: typed contracts, offline fixtures, deterministic traces, safety gates, and gated real-service integrations.

End-to-end intelligent patrol robot combining ROS Noetic, Spring Boot, Vue 3, MySQL, Kinect2, and YOLO for real-world campus security scenarios.

• Led a five-person course team as project lead and primary engineer, owning architecture, core implementation, cross-stack debugging, real-robot deployment, and acceptance.
• Built live mapping and localization, DB-backed waypoint annotation, multi-stop patrol scheduling, direct navigation, run control, teleoperation safety, and real-time dashboards.
• Integrated person and fire/smoke detection with anomaly popups, map markers, silence controls, active tracking, and patrol resumption.
• Verified through 358 backend tests, 52 frontend spec files, gated ROS builds, staged simulation acceptance, and Ubuntu real-robot testing.

Offline-first, byte-reproducible Agent benchmark and main-loop harness integrating GraphRAG retrieval, Graphiti temporal memory, DoT DAG planning, tool constraints, and SLM/LLM routing.

• Ports-and-adapters design with mock/fake defaults and gated real backends.
• Dependency-aware DAG runtime with validation, repair, concurrency, and hard action enforcement.
• 227 tests passed and 6 gated-real tests skipped in the sanitized public snapshot.

Auditable RAG system mapping Victoria 3 player slang to official mechanics through a reviewed concept graph and dual-source evidence.

• Lexical/exact retrieval plus an optional dual-view BGE-M3 semantic channel.
• Read-only MCP tools, late fusion, source references, and fallback/health guards.
• Public code snapshot contains synthetic tests only; raw community/wiki corpora and generated indexes are not distributed.

DoT-inspired adaptive SLM/LLM routing with subtask decomposition, dependency DAGs, alpha-tree supervision, Qwen3 embeddings, and an AdapterMLP router.

• 4,289 labeled subtasks; validation accuracy 95.57%, macro-F1 0.7772.
• Reports class-aware metrics and a clearly bounded P3 evaluation (n=35).
• Separates CPU tests from optional GPU/model dependencies.

AI agents · RAG / GraphRAG · temporal memory · MCP / JSON-RPC · DAG runtimes · SLM/LLM routing · robotics / ROS · computer vision · evaluation · Python

Portfolio: buaahmjh.com/shaozeyv