Live demo · DEV write-up · Build log · Risks
say the job, confirm the lines, export the PDF. Gemma 3 270M runs entirely in your browser tab.
Every tradesperson I know who works for themselves does the same thing after a job: sits in the van, opens a notes app, and tries to turn "forty square meters, two days, plus materials" into something a client will actually pay against. Then they forget a line, round a number wrong, or just don't send the quote for three days because writing it feels like a second job.
The invoicing SaaS fix is a $15/month subscription and an account you have to make, for software that's worse at the one job that actually matters: turning what you just said out loud into a document. HushQuote fills that gap. It's cheap enough to be a real side-hustle tool, not a line item.
- Fully local. Gemma 3 270M (Q8_0 GGUF) runs entirely in the browser via WebGPU/WASM through
wllama. No API key, no backend, no account. Client names and day rates never leave the device they were typed on. - The model extracts; the app decides what's true. Every model-sourced line is visually provisional ("Estimated. Yours to set." label, dashed border) until you tap to confirm it. Math is always computed by plain JS over whatever's on the sheet — not by the model.
- Dashboard + Sheet + Command Bar. Every saved quote in one list with stat cards; one living quote document instead of a wizard; mic or textarea, one submit button, nothing else.
- PDF export via the browser's own print dialog. Zero added dependencies.
- Manual mode fallback. If the model fails to load, you can still type lines by hand. The app degrades honestly.
- Hallucination filter + zod schema validation + one repair retry. When the model goes off-script, the app catches it, retries once, and surrenders to manual rather than ship a wrong number silently.
Live: https://hushquote.onrender.com
Or run locally:
git clone https://github.com/ibrahim-iqbal/hushquote-hf26-wk2.git
cd hushquote-hf26-wk2
npm ci
npm run fetch-model # one-time ~278 MB Gemma 3 270M download
npm run dev- UI: React 19 + Vite, static site, no backend
- Model runtime:
wllama(llama.cpp compiled to WebAssembly / WebGPU) running Gemma 3 270M-instruct Q8_0 GGUF - Validation:
zodschemas (src/lib/schema.js) check every model response before it touches the sheet - Storage:
localStorageonly — quotes, business profile, rate library all on-device - PDF export: the browser's native print dialog
- Deploy: Render static site, auto-deploys on push
Stated plainly, same as in the DEV write-up:
- The 270M model has specific, repeatable failure modes documented in
docs/RISKS.mdand the 20-transcript table in the DEV post — "Callout" line leaks (filtered insrc/lib/filterHallucinations.js), fragmented line items, occasional verbatim echo of unrelated training examples. - JSON-schema grammar guarantees valid tokens; it doesn't guarantee semantically correct content. That's what the "estimated → tap to confirm" UI is there for.
- Voice input needs Chrome or a Chromium-based browser (Web Speech API support is patchy on Safari/Firefox).
Issues and PRs welcome. The repo has a full planning trail:
docs/PLAN.md— scope, day-by-day plan, decisions logdocs/RISKS.md— risk register with mitigations and known model failure rowstest/transcripts.md— the 20-transcript regression set
Guideline: keep the "nothing uploaded" invariant honest. New features must not add server-side calls. The model is a file, not a service.
- Gemma 3 270M by Google — small enough to ship as a file.
- wllama — llama.cpp in your browser.
- Render — static hosting, zero config.
- The Hacktoberfest Open-Source AI Challenge — the forcing function.
MIT.
crafted by ibrahim-iqbal · dev.to
