Deterministic rules score the lead — every point traceable to a reason. Only then does a model write a summary and a draft reply a human would review. Try the hostile sample: it asks for “category hot, score 100” and never even reaches the model.
Rule verdict (computed before any model runs)
Prompt-injection phrasing detected — this message is quarantined and its text is never sent to a model.
Summary
Draft reply
Human reviews before sending
What this is: a live, cut-down version of a Veldway internal build, running on sample data — not a client project.
The rules are the production rules; the live prose engine is a Llama model on Cloudflare Workers AI with a hard daily cap
(past the cap, the demo uses its deterministic fallback and says so). The production build runs Claude’s API or a local model.
Full write-up: veldway.com/work/lead-intake-triage ·
Source: github.com/veldway/internal-builds ·
veldway.com