~/projects/zakat-triage

Zakat-Eligibility Triage

An AI triage agent for crowdfunding campaigns that does everything except decide.

It reads a campaign, assembles the evidence a zakat determination turns on, quotes the organizer in their own words at exact character offsets, names what is missing or contested, refuses when the text cannot carry a determination, and hands the file to a qualified human reviewer. It never issues a ruling. I built it for LaunchGood's Applied AI Engineer process, August to September 2026.

what holds

model calls: 2, claude-sonnet-5, structured outputs
quotes: verbatim, byte-checked, with character offsets
refusal: deterministic, 4 rules
outcomes: human-only, enforced by Postgres CHECK constraints
precedent: pgvector, shown to reviewers only, never to the model
tests: 442
eval gate: 18 fixtures, 100% citation validity
n8n: 3 workflows: pipeline (webhook + hourly sweep), human-approval follow-up, error pager

Next.jsTypeScriptDrizzlePostgres (Neon)pgvectorVercel AI SDKclaude-sonnet-5n8n

Zakat-Eligibility Triage, the reviewer file for one campaign — The reviewer file. Every finding quotes the organizer in their own words at exact character offsets, and each quote is checked byte for byte against the campaign text before it is allowed into the file.

Zakat-Eligibility Triage, a refusal — A refusal. When the text cannot carry a determination, four deterministic rules stop the pipeline and say which one fired. No model judgment is involved in that decision.

The n8n canvas for the triage pipeline — The n8n canvas. A webhook and a scheduled sweep feed the pipeline, a second workflow chases the human approval, and a third pages me on error. n8n owns the clock and the megaphone, the code owns the rules. This capture predates two small changes: the sweep now runs hourly rather than every fifteen minutes, and the test count is 442.

why it never decides

I drew the line where the stakes are. A zakat determination is a religious ruling with money behind it, so the agent prepares the file and a qualified human signs it. Everything in the build serves that line. The quotes are checked byte for byte so a reviewer can trust what they read. The refusal is a rule table rather than a judgment call, so it fires the same way every time. The database will not store an outcome the model wrote. And the precedent search is shown to the reviewer, never fed back to the model, so the past informs the person and not the machine.

› initializing interface

› mounting modules

● online

Zakat-Eligibility Triage

An AI triage agent for crowdfunding campaigns that does everything except decide.