# Letter Decoder: Your AI Friend for Scary Official Letters (Runs 100% Offline)

> Source: <https://dev.to/rambolc/letter-decoder-your-ai-friend-for-scary-official-letters-runs-100-offline-2fji>
> Published: 2026-10-18 18:11:47+00:00

*This is a submission for the [Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle](https://dev.to/challenges/hacktoberfest-week2-2026-10-12)*

**[Letter Decoder](https://letter-decoder.vercel.app)** — a local-first, open-source app that reads your official letters and explains them in plain language.

Tax notices. Fines. Insurance rejections. Bank letters. Landlord warnings. Drop one in as a photo, PDF, screenshot, or pasted text, and get back an **answer card**:

`.ics` calendar export
Every date and amount links back to the **exact line of the original letter** it came from. If the model can't verify a claim against the extracted facts, it says so — no confident hallucinations.

**Who would pay for this?**  Anyone who has ever frozen looking at a government envelope, hired a solicitor at £150/hour just to be told "this is a routine payment reminder", or simply had a letter in their second language. The realistic addressable market is anyone who receives official correspondence — which is everyone.

**The letter never leaves your device.** No accounts, no cloud, no telemetry. That's the product differentiator and the privacy guarantee rolled into one.

|  | URL | Notes | 
|---|---|---|
| **Frontend** | [https://letter-decoder.vercel.app](https://letter-decoder.vercel.app) | React 19 app, deployed on Vercel | 
| **Public API** | [https://letter-decoder-api-qvxs.onrender.com](https://letter-decoder-api-qvxs.onrender.com) | FastAPI — runs in `ENGINE_MODE=mock` , replays 14 canonical examples through the full SSE pipeline with realistic delays | 

For real decoding with a local model, clone the repo and run with Ollama (see below). The public endpoint intentionally uses mock mode so no real letters are ever uploaded to a server.

"A patient friend read it for me, and now I know what to do."

**A local-first, open-source app that explains official letters in plain language.**

Tax notices, fines, insurance rejections, bank letters, landlord notices — drop one in as a
photo, PDF, screenshot, or pasted text, and get back an **answer card**: a one-line verdict, a
deadline countdown, severity, three concrete next steps, a document checklist, and an editable
draft reply. Every date and amount links back to the exact line of the original letter it came
from.

No cloud calls. No telemetry. No accounts. **The letter never leaves your device.**

| URL | What it is |  | 
|---|---|---|
| **Frontend** | [https://letter-decoder.vercel.app](https://letter-decoder.vercel.app) | React app (Vercel) | 
| **API** | [https://letter-decoder-api-qvxs.onrender.com](https://letter-decoder-api-qvxs.onrender.com) | FastAPI engine (Render) | 

The public API runs `ENGINE_MODE=mock`, replaying canonical examples through the same
endpoints and SSE event streams with realistic delays. For real decoding with an open-weight
model, run it locally…

Quick start with a real model:

```
# Terminal 1 — start the inference server
ollama serve
ollama pull qwen2.5:7b

# Terminal 2 — start the app
git clone https://github.com/RAMBO-LC/letter-decoder.git
cd letter-decoder
uv sync --all-extras
uv run python -m api.main
```

Open `http://127.0.0.1:8000` — everything runs fully offline from here.

The core design principle is: **code finds facts, the model explains them.**

```
[Input] → Preprocess → Extract Text → Deterministic Facts → Classify → Notice Pack → LLM Explain → Verify → Answer
```

`engine/preprocess.py`) — OpenCV deskew and denoise for photos/scans.` engine/extract_text.py`) — native PDF text extraction, falling back to Tesseract OCR.`engine/deterministic_extract.py`) — regex + `dateparser` pull every date, amount, and reference number from the raw text. The LLM `engine/classifier.py`) — keyword and embedding matching identifies the notice type and loads the right `notices/`).` engine/llm_constrained.py`) — `engine/verifier.py`) — every factual claim in the LLM output is checked against the deterministically extracted facts. A fact the model invented that wasn't in the source text fails verification — you get a clean error, not a confident lie.
| Layer | Tool | License | 
|---|---|---|
| Local inference | [Ollama](https://ollama.ai) | MIT | 
| Default model | [Qwen 2.5 7B](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct) | Apache 2.0 | 
| OCR | [Tesseract](https://github.com/tesseract-ocr/tesseract) | Apache 2.0 | 
| Image processing | [OpenCV](https://opencv.org) | Apache 2.0 | 
| API framework | [FastAPI](https://fastapi.tiangolo.com) | MIT | 
| Frontend | [React 19](https://react.dev) | MIT | 

Every layer is open source. You can swap `qwen2.5:7b` for Gemma, Llama, Mistral, or a fine-tuned checkpoint by changing one environment variable.

Most contributions to this project are YAML files, not Python PRs. Adding support for a new country is:

```
# notices/fr/tax_reminder.yaml
name: "Avis de rappel fiscal"
keywords: ["Trésor public", "DGFIP", "avis d'imposition"]
severity_default: act_soon
mock_simple: "This is a French tax reminder."
```

That's it. CI validates it against the JSON Schema automatically.

A letter contains income, legal status, financial history, and sometimes your immigration situation. Explaining it with a closed API means uploading it.

There is no enterprise tier, privacy toggle, or data-processing agreement that un-does the fact that your letter travelled to someone else's data center. With an open local model:

`tests/test_local_only.py` fails the CI build if Open weights made this project's privacy guarantee credible. A closed API could promise the same guarantee; open weights prove it.