Inspiration
In Switzerland, a tenant struggles with a German rental contract, in Zurich a SME owner can't parse a tax form in French, and in Chakwal, Pakistan, my mother can't read a land record in English. All face same problem: they must upload sensitive CNIC, contracts to US closed AIs, losing data sovereignty. Switzerland built Apertus as 100% open, sovereign LLM - we asked: can Apertus explain civic docs in your own dialect, fully offline?
What it does
Nyaya Saathi is a privacy-first assistant powered by Apertus.
- Upload any civic/legal PDF: Swiss tenancy contract, tax notice, Pakistan domicile, SME compliance form
- Choose your dialect: Swiss-German, French, Italian, English, Urdu, Punjabi & more
- Get simple, point-wise explanation in your language, with key risks highlighted
- 100% offline via Ollama - no data leaves your laptop/phone. True sovereignty.
From a farmer in Chakwal to a shop owner in Bern - same model, same trust.
How we built it
- Model: Apertus-8B-Instruct running locally via Ollama, quantized to 4-bit for low-cost devices (Toshiba laptop tested)
- Pipeline: PDF -> PyMuPDF text extraction -> Prompt with dialect instruction -> Apertus -> Simple explanation + JSON of key clauses
- Frontend: Gradio for offline demo, supports 6 languages
- Localization: Few-shot prompting for dialects, not full fine-tune, to stay lightweight and open
Challenges we ran into
- Running 8B model on old hardware - solved with Q4 quantization
- Swiss-German dialect has little training data - used few-shot examples from locals
- PDF parsing for Urdu nastaliq script - added OCR fallback
Accomplishments that we're proud of
Proved Apertus goes beyond Swiss borders - from Alps to Punjab, sovereign AI can serve 230M Urdu/Punjabi speakers offline, same as Swiss dialects. Aligned with Apertus mission of open, inclusive, sovereign AI for public good.
What's next
- Add voice: offline Urdu/Punjabi TTS like our Dawat AI project
- SME compliance agent: auto-check contracts + monthly subscription model
- Expand to 20 dialects, publish open dataset of civic doc explanations
Built With
- apertus
- civic-tech
- gradio
- multilingual
- ollama
- open-source
- pymupdf
- python
- sovereign-ai
- swiss-german
- urdu
Log in or sign up for Devpost to join the conversation.