AI Operations
I work with AI tools daily — not as a researcher, but as a practitioner who deploys, optimizes, and automates with them. I bring the same infrastructure discipline to AI operations that I apply to servers and clusters: measure everything, automate what repeats, and cut waste ruthlessly.
What I Do
LLM Context Management
Large language models are expensive to run. Every conversation round-trips the full context — a 10-turn session on a 7,000-line codebase can burn nearly a million tokens, most of it redundant.
I built a context architecture that replaces brute-force loading with lean pointers and indexed search. Operational knowledge lives in Git repos, referenced by compact memory files the agent loads at conversation start. When detail is needed, an MCP tool queries a MeiliSearch index across all 38 repos in under 50ms — with fuzzy matching and relevance ranking that sequential grep cannot provide. The result: an 18% reduction in always-loaded context with a vastly larger effective knowledge base.
This matters for any organization running LLM workloads at scale. The same pattern — pointer-based memory, indexed retrieval, MCP tooling — applies wherever agents need access to large knowledge bases without burning context on content they may never read.
AI-Assisted Workflows
Daily Claude Code usage for infrastructure management, code review, documentation, and automation. Not toy prompts — production work: deploying Kubernetes services, debugging mail delivery, writing monitoring alert rules, managing DNS zones.
I've built and archived over 1,000 AI conversations across Claude, DeepSeek, and Mistral. That volume teaches you what works and what doesn't — how to prompt effectively, when to use which model, and where AI falls short.
Speech-to-Text Pipeline
Automated voicemail transcription using whisper.cpp running locally on private infrastructure. IMAP polling retrieves audio, whisper.cpp transcribes it, and ntfy delivers push notifications with the transcript. No cloud APIs, no per-minute billing, no data leaving the server.
Multi-Model Access
CLI-based access to 26+ models via OpenRouter, with shortcuts for quick queries, data analysis via piped stdin, and model selection by task. Different models for different jobs — not every question needs the most expensive answer.
What I Bring to a Team
Most people who understand AI can't deploy infrastructure. Most people who manage infrastructure haven't used AI tools in production. I do both, every day.
- Practitioner first — I use these tools to do real work, not demos
- Cost-conscious — I've optimized my own token spend and can do the same for yours
- Infrastructure-native — I deploy AI tools on infrastructure I build and maintain
- Can train others — seven years of teaching means I can bring your team up to speed, but I'd rather be building