Software engineer building reliable AI systems at the intersection of agent engineering, local-first applications, and embodied AI.

My work focuses on turning model capabilities into explicit, testable software: structured contracts, bounded workflows, durable state, provenance, evaluation, and human control. The two projects below are the foundation of that development path.

A local-first system that records authorized meeting audio, produces transcripts with local speech recognition, and derives auditable meeting intelligence grounded in source segments.

What it demonstrates

• FastAPI control panel and local system-audio capture on macOS and Linux
• Local ASR with faster-whisper and low-resource defaults
• Checkpointed LangGraph workflows with structured, provenance-aware outputs
• Optional local intelligence through LangChain and Ollama
• Meeting-scoped, deny-by-default MCP resources and tools
• SQLite WAL persistence, consent evidence, revision invalidation, and deletion receipts
• One-command startup, automated tests, strict typing, and cross-platform CI

Python · FastAPI · faster-whisper · LangGraph · LangChain · Ollama · MCP · SQLite

A simplified architectural reproduction of Harness VLA on EmbodiedBench's EB-Manipulation environment, built to study reliable manipulation through a planner, a fixed primitive library, memory, simulator feedback, and auditable traces.

What it demonstrates

• Closed-loop planning through one validated JSON primitive per turn
• Separation between analytical motion and contact-oriented vla_act
• Global and task-specific memory experiments with evidence gates
• CoppeliaSim/PyRep integration and EB-Manipulation evaluation
• Incremental traces, postcondition checks, failure analysis, and reproducible run reports
• Paper-grounded implementation milestones and explicit beta limitations
• Specialist development agents for paper research, implementation, simulator diagnosis, and trace analysis

The current repository validates the architecture and observability of the harness. It does not claim a complete reproduction of the paper: the real frozen visual VLA backend and full experimental protocol remain roadmap work.

Python · Embodied AI · VLA · EmbodiedBench · CoppeliaSim · PyRep · Ollama · Agent Evaluation

These projects explore a shared question: how do we make AI systems dependable beyond a successful model call?

I am developing that answer around a few recurring principles:

• explicit state and typed boundaries;
• deterministic infrastructure around probabilistic models;
• local and privacy-conscious execution where practical;
• evidence-linked outputs instead of unsupported claims;
• measurable evaluation, failure traces, and honest scope;
• small validated iterations backed by tests and documentation.

• Reliable agent orchestration and memory
• Local speech and meeting intelligence
• Vision-language-action systems and robotic manipulation
• Evaluation harnesses, provenance, and observability
• Python backend architecture and developer tooling

The repositories above are active engineering notebooks as well as portfolio projects: each decision is expected to become code, a test, an experiment, or a documented limitation.