AI infrastructure engineer focused on operationally reliable agent systems.
I build local-first AI tooling for cost visibility, evaluation discipline, workflow governance, MCP operations, and multi-agent orchestration. My background is 17 years in manufacturing and logistics operations, which shows up in the engineering style: standard work, observability, error-proofing, decision logs, and measurable validation.
| Area | Repo | Why It Matters |
|---|---|---|
| AI operating environment | animus | Multi-package AI system with memory, workflow orchestration, quality gates, and autonomous improvement loops. |
| Cost observability | ai-spend | Terminal-native spend aggregation across Anthropic, OpenAI, and OpenRouter without proxying requests. |
| MCP operations | mcp-manager | CLI for discovering, validating, and syncing MCP servers across agentic IDEs. |
| Workflow governance | agent-lint | Static analysis for agent workflow YAML: cost exposure, retry limits, schemas, and failure-prone patterns. |
| AI engineering method | the-human-stack | Evidence-graded engineering reference for production AI systems. |
| Evaluation practice | arete-evals | LLM eval suites and run records behind the methodology. |
- Systems engineering: multi-package Python systems, FastAPI, SQLite/PostgreSQL, Pydantic, Rust/PyO3 boundaries, GitHub Actions.
- AI operations: evaluation pipelines, cost controls, workflow checkpoints, quality gates, MCP lifecycle management.
- Security posture: local-first defaults, secret hygiene, permission-prompt auditing, CodeQL/security workflows where applicable.
- Documentation: decision logs, evidence grading, architecture docs, install paths, demo GIFs, and runnable examples.
Personal AI operating environment with local-first control, evidence-graded maturity, bitemporal memory, workflow orchestration, and autonomous improvement proposals.
Recruiter signal: large system design, package boundaries, test discipline, docs, CI hardening, operational framing.
Start here: README - Architecture - Roadmap
CLI for aggregating AI API costs across providers without adding a proxy, SDK wrapper, or hosted dashboard.
Recruiter signal: practical developer tooling, provider adapters, local credential handling, terminal UX, billing API integration.
Start here: README - pip install ai-spend
MCP server lifecycle manager for Claude Code, Cursor, Windsurf, and related agentic IDE workflows.
Recruiter signal: emerging AI tooling infrastructure, config normalization, health checks, security-aware developer workflows.
Start here: README - pip install arete-mcp
Static analyzer for agent workflow configs. Detects unbounded retries, missing cost guards, hard-coded secrets, missing schemas, and fragile parallel orchestration.
Recruiter signal: AI governance translated into executable checks.
Start here: README - pip install agentlinter
Most AI projects fail for operational reasons before they fail for model-capability reasons: unbounded cost, unclear validation, brittle workflows, and no recovery path after partial failure. My work focuses on making those risks visible and controllable with normal engineering mechanisms.
That operating model comes from manufacturing and logistics: define standard work, instrument the process, stop defects early, keep evidence, and improve the system continuously.
- The Human Stack - evidence-graded methodology for operating AI systems.
- Substack - writing on AI operations, evaluation, local-first tooling, and manufacturing discipline applied to software.
- Decision Logging as Operational Memory - ADL practice for keeping architecture decisions inspectable.