README.md

metadata

title: Bureaucat
emoji: 🐱
colorFrom: pink
colorTo: yellow
sdk: gradio
sdk_version: 6.16.0
python_version: '3.12'
app_file: app.py
pinned: false
license: apache-2.0
short_description: The cat that reads scary Swedish letters for you
tags:
  - track:backyard
  - achievement:offgrid
  - achievement:offbrand
  - achievement:fieldnotes
  - build-small-hackathon
  - backyard-ai
  - off-the-grid
  - off-brand
  - field-notes
  - zerogpu
  - vision-language-model
  - sweden
models:
  - Qwen/Qwen3-VL-8B-Instruct

🔗 Submission: Demo video & launch post on LinkedIn — the demo video is embedded in this post.

🐱 Bureaucat

The cat that reads scary Swedish letters so you don't have to.

You moved to Sweden. One day a thick envelope arrives from Skatteverket. Or Försäkringskassan. Or — heart rate rising — Migrationsverket. It's in formal bureaucratic Swedish, it mentions an amount and a date, and you have no idea if it means "FYI, everything is fine" or "pay 15,600 kr in 30 days or lose your housing allowance."

Bureaucat reads the letter for you and tells you, at a glance and accurately:

• How worried to be — a 1–5 severity verdict driving a Panic Meter and a reacting cat-civil-servant mascot
• The short version — what the letter says, in plain English
• Why you got it — which authority, and what triggered it
• What you need to do — a checklist of required actions
• Deadlines & money — every deadline, amount, and reference number quoted verbatim from the letter, never invented

🏡 Backyard AI: built for a real person — me

I'm an expat in Sweden. Every letter in the example gallery is a type of letter that has actually landed in my mailbox: tax decisions (slutskattebesked), benefit completion requests (komplettering), residence-permit document requests, CSN repayment demands, vårdcentral appointment summons. The first user test was me, on my own mail. The app is stateless by design — nothing is stored, nothing leaves the Space.

🤏 Small model, honest fit

One model does everything — OCR, reasoning, and explanation in a single pass:

Model	Qwen/Qwen3-VL-8B-Instruct (~9B params — well under the 32B cap)
Decoding	Greedy (do_sample=False) — deterministic output for an accuracy-critical tool
Inference	Entirely in this Space on ZeroGPU. No cloud APIs.

The model was chosen by a bake-off, not vibes: Qwen2.5-VL-7B vs Qwen3-VL-8B on a gold set of five synthetic Swedish authority letters with an automated evaluation harness (eval/run_eval.py). The smaller 7B model dropped reference numbers (3/5 pass); the 8B passed 5/5 with zero invented values and 100% recall — so per "smallest model that passes," the 8B won.

🛡️ The anti-hallucination contract

Inventing a deadline is the single worst failure this app can have. So:

1. The model first transcribes the letter, then extracts — every value in "Deadlines & money" must be a verbatim substring of its own transcription, checked by a pure-Python grounding pass (no second model call).
2. If a value isn't in the letter, the answer is "None found." — never a guess.
3. The eval gate (zero invented values, 100% recall, severity parseable, beginner-mode invariance) runs on every gold letter, plus adversarial fixtures: a blurry photo, a grocery receipt, and a non-Swedish letter must all be refused, not analyzed.

👇 How to use it

Upload a saved file — a photo (JPEG/PNG) or a PDF of your letter, multi-page supported. Bureaucat does not use your camera or take live photos; snap the letter with your phone first, then upload the image or PDF here. One upload uses one short ZeroGPU call.

The example gallery at the bottom of the app contains pre-computed analyses — tap any of the five letters and the full result loads instantly at zero GPU cost, no upload needed.

Why synthetic examples? Real authority letters are full of personal data (names, personnummer, amounts) and can't be published — and the official "example" documents authorities publish are info sheets without the personalized deadlines/amounts that make a letter scary. So the gallery letters are faithful synthetic recreations of the five letter types, with realistic Swedish layouts, diacritics, and values — built so nobody's real mail ends up in a public repo. Every letter shown in the app and in the demo video is one of these synthetic mock-ups — I tested Bureaucat privately on my own real Swedish mail, which never gets published.

🏅 Badges claimed

• Off the Grid — no cloud APIs; the 9B VLM runs entirely in this Space
• Off-Brand — custom frontend via gr.Server
• Field Notes — build report: what greedy decoding revealed that sampling had masked

⚖️ Privacy & disclaimer

• Stateless: letters are processed in-memory and never stored. Still — redact your personnummer before uploading if you can.
• Not legal advice. Bureaucat helps you understand a letter; always verify deadlines against the original and contact the authority if unsure.

---

Built by an expat, for expats, with one small cat-shaped model. 🐾