I'm an AI Engineer building the infrastructure and products that turn intelligent models into usable systems. My work spans agent reliability, orchestration, local inference, applied machine learning, simulation, and decision intelligence.
I care about what happens beyond the demo: architecture, evaluation, operational boundaries, and whether the product solves a real problem.
|
Reeprov ↗ Local-first debugging infrastructure for AI agents: capture, replay, causal comparison, evaluations, and regression gates. Agent reliability · TypeScript + Python SDKs · Public beta |
OpsTwin ↗ Decision lab comparing operational changes through paired, repeated simulation and uncertainty-aware evidence. Next.js · FastAPI · SimPy · 430 reported passing tests · Live demo |
|
BI Notebook Lab ↗ Browser-based DAX learning lab with an expression engine, semantic models, filter propagation, and executable grading. TypeScript · 1,024 reported passing tests · 82/83 hand-verified DAX cases · Live demo |
Relay ↗ A source-aware marketing analytics workspace for Meta Ads, Google Ads, and Shopify CSV exports. Next.js · Data quality · 304 reported passing tests · Protected beta |
|
Open VS Code Agent ↗ A coding-agent system optimized for a local 7B open-weight model, with strict tool boundaries and reproducible evaluation. Ollama · Local inference · 37.5% → 75% success on a 24-task read-only benchmark · Public architecture edition |
DemandOS ↗ Demand forecasting and inventory-risk planning with leakage-aware feature engineering, backtesting, and human-reviewed reorder proposals. Python · Applied ML · Synthetic-data portfolio MVP · Live demo |
- AAO / Applied AI Operations — a private, reusable foundation for governed AI workflows, execution, approvals, and adapter portability.
- IRIS OS — an attention-first personal AI operating system with bounded specialist agents. The public repository documents the architecture; external side effects are not production-enabled.
- HALO Control — a local AI control plane for workload orchestration and model routing. The public repository is an architecture edition, and AMD Halo hardware verification remains pending.
Other experiments: DataBrief AI · SearchSignal · Website Audit Agent · Overflow iOS
Agent orchestration · MCP · Local inference / Ollama · Evaluation harnesses · Playwright · Pytest · Vitest · SimPy · DAX
- Lemonade — backend configuration validation — merged contribution rejecting invalid backend-specific configuration keys.
- OpenLIT — LangGraph memory connector — contribution proposing a memory connector with CRUD/search, authentication, and integration tests.
- OpenJudge — empty-string grader edge case — contribution addressing degenerate empty-string matches.



