Quotengine

Overview

An estimate engine for dental practices. A patient describes the treatment they need, optionally with a panoramic x-ray, and gets an indicative quote.

Stack

TypeScript, Hono, Cloudflare Workers, D1, R2, Claude (vision and tool use), zod, Vitest

What I built

• Guarded multimodal intake: Claude reads the patient's message and an optional panoramic x-ray, and can only answer through a forced tool schema that picks items from the practice's catalog.
• A deterministic pricing engine in TypeScript: per-practice catalogs, rules, modifiers, tax and markup, all validated with zod before they load.
• A Hono API on Cloudflare Workers with leads in D1 (versioned SQL migrations), radiographs in private R2 storage and an OpenAPI contract.
• Health-data privacy: consent is recorded with the SHA-256 of the exact text accepted, and an x-ray is stored only with a separate opt-in, otherwise it is processed in memory.
• 333 tests across 44 files running in CI, plus an evaluation harness on open-access radiographs.
• Docs a reviewer can follow: four architecture decision records, a threat model, a model card and a go-live privacy checklist.

Key decisions

• The model never does arithmetic. It returns catalog IDs through a forced tool schema, and code computes every price.
• Uncertain input fails closed: a poor x-ray or a low-confidence reading returns a consultation request instead of a guessed price.
• No accuracy claims without proof: the model card states what has not been measured yet on a dentist-labelled dataset.
• Every estimate is indicative and needs in-person clinical confirmation.