AI application engineering learner building reliable, testable agent systems.

I care about the part after the demo: bounded tool use, evidence, permissions, failure handling, human confirmation, and a clear boundary between model suggestions and business writes.

Trusted e-commerce AI after-sales workflows built on top of macrozheng/mall.

Natural-language goal
  → FastAPI Agent Runtime
  → versioned Skills + read-only fact investigation
  → RAG evidence / Context Pack / replanning
  → ActionProposal + explicit confirmation
  → Java authority: ownership, state machine, idempotency, transaction, final write
  → MySQL Outbox / RabbitMQ event

The project is a local, synthetic-data engineering demo. It is not a production SaaS and does not claim real payment, warehouse, logistics, repair, production-SLA, or real-user accuracy.

Evidence and boundaries:

• README and local demo
• Architecture
• Release gate summary
• Resume fact pack

• Agent Runtime: task planning, versioned Skill Catalog, bounded execution, replanning, and safe terminal states
• Context Engineering: artifact projections, Context Pack compression, owner-scoped memory, TTL, and recovery boundaries
• RAG Engineering: chunk metadata, version filters, Dense retrieval, BM25/RRF and reranker comparisons, grounding and abstention
• Reliable execution: ActionProposal, confirmation gates, Java-authoritative writes, idempotency, Outbox/RabbitMQ, and failure stop
• Evaluation: deterministic contract gates, synthetic live-model tests, trace allow-lists, red-team cases, and stale-report detection

Java · Spring Boot · MyBatis · FastAPI · Pydantic · LangGraph · Vue 3 · TypeScript · Redis · MySQL · MongoDB · RabbitMQ · Docker Compose · Chroma · BM25 · MCP · pytest · Maven · GitHub Actions

• Facts, permissions, state transitions, idempotency, and final writes belong to the domain service.
• The model can propose a plan or a read-only next step; it cannot invent tools or write business data directly.
• Evidence is versioned and inspectable. When evidence or a dependency is unavailable, the system stops safely.
• Test counts are always tied to a command, commit, fixture hash, and execution mode.
• Upstream code and AI-assisted work are disclosed clearly.

I am preparing for AI application engineering and Agent engineering roles, with an emphasis on turning prototypes into systems that can be explained, tested, and operated.

Build → measure → inspect failures → tighten the boundary → document the evidence