AI Agent Developer building reliable, auditable systems that connect models, tools, data, and human decisions.

I am an AI agent developer focused on turning model capabilities into dependable software. My work connects agent orchestration, tools, data, human decisions, and verifiable delivery.

• Building reliable agent systems with bounded execution, explicit state, recovery paths, and controlled side effects.
• Debugging from contracts and regressions across async workflows, MCP tools, streaming state, accessibility, and interoperability.
• Producing engineering evidence through deterministic evaluation, observability, automated tests, and CI.
• Contributing upstream with focused fixes, transparent validation, and careful review follow-up.
• Drawing on a practical engineering background in Python, TypeScript, C#, MATLAB, desktop utilities, signal processing, and visualization.

An auditable operations agent for inventory exceptions, equipment alerts, and human-approved work orders.

Background: Operational exceptions need more than a model response. The system must investigate evidence, stop before side effects, bind approval to current facts, recover safely, and leave an audit trail.

Completed work:

• Built one bounded LangGraph runtime for inventory, equipment, and blocked-task scenarios.
• Added FastMCP tools, policy checks, human approval checkpoints, idempotent writes, restart recovery, and read-back verification.
• Integrated PostgreSQL, pgvector, Redis, OpenTelemetry, React, Docker, deterministic evaluation, and CI-backed release evidence.
• Kept the model advisory while deterministic code owns write authorization and safety boundaries.

A policy-aware field mission orchestration agent for multi-stop business travel, expense constraints, and event-driven replanning.

Background: Field work combines fixed appointment windows, intercity and local transport, accommodation, meals, reimbursement rules, and disruptions. FieldPilot turns those constraints into explainable plans instead of relying on an unconstrained model response.

Completed work:

• Built a typed PydanticAI intake boundary with persisted missions, revisions, events, provider snapshots, and agent-run evidence.
• Added bounded candidate search, a deterministic policy engine, and an independent verifier for time, cost, task coverage, and meal constraints.
• Integrated asynchronous Amap route and nearby-POI adapters with caching, request budgets, in-flight deduplication, and explicit fixture fallback.
• Delivered event-driven replanning with revision diffs, a Vue workbench, Alembic migrations, Docker configuration, CI, and 42 backend tests.

• Serial Port Assistant and KLS Serial Utility: C# Windows tools for serial communication and protocol testing.
• MATLAB Imaging Tools: sensor frame parsing, signal imaging, trajectories, and 3D visualization.
• PDF Report Generator: C# reports containing text, images, JSON, tables, watermarks, and page numbers.

I contribute focused, regression-tested fixes to established AI agent and developer-tool projects. Personal-repository merges are excluded from the upstream count.

My contribution work also covers async stream cleanup, token-usage normalization, human-in-the-loop continuation, cache migrations, A2A interoperability, and agent UI state management.

I am open to conversations about reliable AI agents, MCP integrations, human-approved automation, and focused open-source collaboration.