Projects

Everything here exists and runs, or I say plainly where it doesn't yet. I built these because I handle the invoicing and adjuster responses for hundreds of open water-damage claims at once, mostly on my own. Each one has a longer write-up in my working notes.

ClaimForge

ClaimForge turns an approved scope sheet into a final invoice and a review package. It reads the scope sheet, builds the invoice in Invoice Ninja, keeps my notes separate, and records whether each price came from the catalog, the contract rate or a manual decision. The quantities from the field are never edited in place; any proposed change sits beside them for review. Its test runner reports PASS, SKIP with a reason, or FAIL, so a missing input never passes as a failure or a success. The source is integrated and tested; acceptance of the installed release is still open. The code is private. Why I built ClaimForge.

Aries

Aries is an office assistant that runs on its own Mac mini. It is Claude Code running headless, driven by about 20 scheduled jobs, each following a written recipe and keeping its state in plain files. Every call goes through one wrapper with its own capped budget and a spend log. Each morning it sends me a briefing. What Aries actually does.

The claim graph

I needed exact answers to questions like "which claims does this adjuster handle" and "what did this carrier pay before." An AI-built graph could not answer them reliably, so I built one from records with plain code: the client registry, the document library, mail headers and the invoice register. It rebuilds every hour. On a 25-question test it got 21 right, and lookups take under a third of a second. It does not model line-item facts yet. How the claim graph works.

Three local-model machines

Three AMD Strix Halo machines run open models locally through the Lemonade server: a coding model, a vision model that reads job photos and meter readings, and a large model used as a checker. They take the bulk reading work, where speed matters less than cost. A box reads about one job photo every three to four seconds, so a backlog of 10,000 photos is an overnight job. What actually works on Strix Halo.

The document library

Contracts, invoices, dry logs, carrier estimates and correspondence now live in one library on a network drive, with a record of where each file came from. Before it existed, eight separate stores held the same 1,897 files three times over with nothing cross-referenced. The local models and the claim graph read from this one tree.

Stack and open source

The back office runs on Cloudflare and Supabase, with pg_cron as the only scheduler: the stack. Public code and how I use Claude Code and Codex are on faroughdarakhshan.org.