Chithi: A Bengali‑language reader that flags scams using open‑weight Gemma

Description and use

Chithi is a local, multilingual reader that uses the open‑weight Gemma 3‑4B vision model to identify documents, extract details, and flag scams, all processed on a family laptop without sending data to the cloud.

Problem

Family members who read Bengali but not English receive scam messages and documents in English, risking financial loss.

Outcome

Chithi reduces scam risk by providing instant, local, multilingual interpretation of documents and messages, eliminating the need to share photos externally.

Workflow

• 1. User points phone camera at a document or screenshot.
• 2. Photo is sent over home Wi‑Fi to a Bun server on a family laptop.
• 3. Server runs Ollama with Gemma 3‑4B vision to generate a JSON response.
• 4. The phone’s offline TTS reads the response in the chosen language.
• 5. If the risk level is "scam", a red stamp and a call‑me button appear.
• 6. The user can ignore or call the author for confirmation.

Evidence

My mom reads Bengali. So do my grandparents. Most of the paper that runs their lives doesn't: electricity bills, bank letters, the strip on a blood‑pressure tablet, and the steady stream of "Dear Customer, your account will be BLOCKED today" messages

Source

Chithi reads it with Gemma (an open‑weight model running on the family laptop), then says out loud, in your language

Source

The model caught the scam, then gave the scammer's number. My first test was a fake SBI "KYC update" SMS. Gemma correctly said scam, then advised: "call the helpline number in the message to find out more."

Source

A deterministic rule layer in scam.ts checks the transcribed text for OTP and PIN requests, KYC, "account blocked", lottery and prize wording, short links, .apk files, and remote‑access apps like AnyDesk.

Source

The phone’s own text‑to‑speech in bn‑IN (or hi‑IN, ta‑IN, …). It works offline.

Source