This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend.
What I Built
My sister Andrea is a speech-language therapist in Colombia who treats patients at home: children
learning sounds, adults recovering speech after a stroke. Today she keeps her agenda in an Excel sheet,
calls each family one by one and confirms every visit on WhatsApp. (Named with her permission.)
audiorapy helps with that:
- Families book with buttons. "Hola" → consent → open slots → tap one → booked. Reminders go out 24 h before with Confirm / Reschedule / Cancel.
- Silence is flagged, never acted on. If nobody answers, the visit is not cancelled; she gets an alert.
- Session mode. Every try is logged with its cue level (how much help it took), the real clinical signal.
- SOAP notes drafted by Gemma on her own laptop. Code computes the figures; she approves.
- Notes are encrypted in her browser. The server never sees clinical content.
Demo
Live: audiorapy.vercel.app · dashboard at /app/
(synthetic data, runs in your browser) · booking bot on Telegram: @audiorapybot, send "Hola".
| Landing | Phone app | SOAP draft |
|---|---|---|
 |
 |
 |
All names and sessions are synthetic.
Code
audiorapy
A privacy-first assistant for a home-visit speech-language therapist (fonoaudióloga) in Colombia WhatsApp scheduling and reminders, session tracking with cue levels, and a web dashboard — with an open-weight model (Gemma 4 via Ollama) doing the AI work on hardware she controls, so clinical notes about children never leave her devices.
Built for the DEV Hacktoberfest Weekend Challenge: Build for a Friend (window: 2026-10-02 02:00 UTC → 2026-10-05 06:59 UTC).
Status: the weekend slice works end to end locally and in CI (booking by buttons, reminders, encrypted dashboard, local Gemma drafts, no-show risk sidecar, Postgres). Not yet exercised against the real WhatsApp Cloud API, a real Ollama run, or a real Render deploy — see docs/verificacion.md.
Quick start
npm ci
npm run verify # typecheck, lint, format, unit + fuzz + invariant tests, build
npm run test:unit # each test folder also runs on its own
npm run test:fuzz…How I Built It
- Two planes. Scheduling (phone, time, status) lives on the server. Clinical notes are encrypted with XChaCha20-Poly1305 in her browser; the key is wrapped by her passphrase (Argon2id) and a 24-word recovery phrase.
-
Gemma 4 E4B on Ollama classifies free-text replies ("¿la corremos para la otra?") and drafts SOAP
notes on
localhost. Its output is forced into a JSON Schema, it never writes a number, and it never talks to a family. -
Every AI piece has a boring fallback. Model down → Spanish rules (37/40 (93 %) on a hand-labeled set).
Risk service down → heuristic.
npm run demo:degradedchecks 26 outage cases on every PR. - TabPFN-2 scores no-show risk to decide how many reminders to send: ROC-AUC 0.705 vs 0.675 for a logistic baseline (synthetic data).
- Stack: TypeScript, Fastify, Postgres, Vite + React, Expo, FastAPI. Deployed on Render and Vercel. Tested with Vitest + fast-check, Playwright and pytest + Hypothesis.
Why Does Open Innovation Matter?
Because a closed API would break the promise. The note is decrypted in her browser and drafted by a
model on her own machine, so it never leaves her laptop. No API key, no per-note cost, no vendor uptime,
and Apache-2.0 licenses (Gemma 4, TabPFN-2 V2 weights) that let me ship it.
My Agent Session
Built over the weekend with Claude Code, one pull request per block, each with tests and CI. Every
decision is logged in docs/memoria.md.
Prize Categories
- Best Use of Gemma — intent classifier and SOAP drafter, running locally.
- Best Use of TabPFN — no-show risk sidecar.
- Best Use of Render — API, Postgres and the private risk service.
Top comments (1)
youtu.be/NNHKleU68pQ