Mike Martin.

The full public web presence for a historic inn, tavern, and events venue — fourteen pages, a complete design system, and the operations plumbing behind it. Designed, built, migrated onto the client’s own hosting with zero downtime, and handed off with documentation the owner can actually use.

• Build engine with mechanical deploy gates — every deploy is linted, probed, and verified before it counts
• Events calendar feeds live from the venue’s own Google Calendar; the owner never edits HTML
• Transactional email with branded guest auto-acknowledgment, provider-scoped DNS auth that survives host moves
• Migration rehearsed end-to-end on scratch hardware first; three independent rollback layers held during cutover
• Lighthouse 100s on accessibility, SEO, and best practices sitewide

My daily engineering practice: frontier models do real labor — writing code, running migrations, building features — and I engineer the harness that makes their work trustworthy: toolchains the models operate through, gates their output must clear, verification that decides what ships. Cynthia below is the deepest expression — an AI platform, built AI-natively.

• Models operate real toolchains — deploys, database surgery, browser QA — never freehand
• Every output clears mechanical gates: lint, probes, contract checks, parity and mutation proofs
• The harness is the craft: work that can’t pass the gate can’t ship

Can a cheap model author spec oracles for real shipping code — without anyone having to trust the model? Built on the published agentic property-based-testing work, which suffers a large invalid-report rate when the model judges its own output. Here an oracle only counts if it passes on the real code and kills mutants of the function it claims to specify: the model proposes, mutants dispose.

• Controlled experiments — leave-one-out across four target repos, pre-committed holdouts, a null result reported as null
• A model-free parser-differential auditor for the SSRF bug class, findings cited to the RFC clause
• Reports on real libraries keep their post-hoc corrections visible — an auditor that can't say "we were wrong" produces noise

A platform that researches a local business, plans from a discovery corpus, writes its own pages, and judges its own output — built as a typed plugin kernel where every LLM concern (extraction, validation, retries, concurrency, durable execution) lives once as a service instead of being re-rolled per feature.

• ~25 services and 100 tools on a search-based disclosure kernel — agent context stays flat as capability count grows
• Every model output crossing a boundary is contract-checked; broken output is structurally unshippable
• Multi-agent swarm execution and durable retries on a Postgres-backed workflow engine
• Migration from the previous architecture proven with parity ledgers and mutation-tested gates

The system that built the Temperance House site — generalized into an engine any site profile can drive. Page builds are code, presets are a round-trip-proven library, and deploys refuse to ship anything that fails the guards.

• Pre-ship shortcode linting and post-deploy escape-leak probes on every push
• Preset library regenerated from the live database and drift-checked — 36/36 round-trip identity
• Written to be operated cold by an AI agent: docs and mechanical checks over tribal knowledge

The production floor for everything above: a Proxmox fleet running the development, inference, and management planes, with cloud LLM traffic consolidated behind one gateway.

• LiteLLM gateway with a custom streaming-cache shim — ~82% cache hit rate on agent workloads
• Durable job workers, uptime monitoring, reverse proxies, and a Tailscale mesh across every device
• Managed like production: inventoried, monitored, and documented