Projects
AI, automation and infrastructure projects that started as real operational needs of my own.
AI Support Operations Copilot
Status: Prototype complete · Not deployed
A human-in-the-loop support operations system that analyzes synthetic support requests, validates structured outputs, identifies security-sensitive content and prepares reviewable response drafts — with a measured LLM layer behind a fail-closed contract.
Current milestone:
- Deterministic baseline + optional LLM engine (Claude / OpenRouter) behind one contract
- Synthetic evaluation snapshot (4 August 2026): correct categories on 33 paraphrased requests — baseline 14/33; four LLM models 23–33/33, with rejected calls counted as failures. These are test results, not production accuracy.
- Independently audited labels, a time-bounded baseline fallback, and guarded LLM reply drafting
- Single-source rule data generating the n8n workflow byte-identically
- 94 automated tests, public CI, native verification on self-hosted n8n 2.14.2
Quiet VPS Watchdog
Status: Retired (August 2026)
A deterministic readiness watchdog for a self-hosted VPS: disk, memory, systemd, Docker health, HTTP, TLS expiry and backup age, deduplicated by a content signature so it alerted once when something broke, stayed silent while it persisted, and reported once when it recovered. No LLM in the check path, no autonomous restarts. Retired with the server when its workloads moved to managed platforms; kept as a record of the design.
Also covered: read-only n8n access for an AI agent through a local MCP bridge, with mutation tools excluded at the client.
ShiftScan
Status: Live demo · In development (Render free tier, may take up to a minute to wake)
A photo of a shift roster becomes an ICS calendar; an optional planner places user-defined activities into the free time. Stateless by construction, LLM output validated row by row, and a rule-based planner that keeps the app fully functional without a model.
Coming next
- Operations Incident Timeline Analyzer