"Rules do not exist for the sake of control. They exist for the sake of compatibility. But what happens when every state has its own track gauge?"
💡 Inspiration
When reviewing the Hack Apertus challenge tracks, I wanted to build something fundamentally useful for society — a system tackling a real problem, while pushing my own engineering boundaries by working hands-on with the Swiss National Supercomputing Centre (CSCS Alps).
Switzerland’s federalist strength is also its greatest data paradox: 26 cantons, 4 national languages, and ~300,000 parliamentary affairs trapped in disparate, idiosyncratic PDF documents.
In the Track 2A (OpenParlData) challenge description, three critical gaps were highlighted: the Format Gap, the Extraction Gap, and the Verification Gap. When rules and formats diverge like railway gauges, democratic transparency grinds to a halt. I set out to prove that in just 7 days, a sovereign, open-weights AI model could bridge these gaps into a single, unified, verifiable standard.
⚙️ What It Does
SEVERIN (Swiss Extraction & Verifiable Entity Representation for Parliamentary Inter-cantonal Networks) is a sovereign AI extraction platform that ingests heterogeneous Swiss parliamentary documents across all 26 cantons and the Federal Assembly:
- Unified Semantic Normalization: Translates chaotic cantonal PDF layouts across 4 languages (DE, FR, IT, RM) into a single, rigorous Pydantic v2 schema (
ParliamentaryAffair) harmonizing 8 affair types. - 100% Verifiable Provenance: Every single extracted fact, author, question, and government reply is tied to a 1-indexed physical PDF page number and an exact verbatim quote.
- Zero Hallucinations by Design: Enforces the "Rule of Absence" — if an affair does not allocate budget credits or does not have an executive response yet, fields return strict
null/[], never invented numbers. - Interactive Full-Stack Platform: Features a split-screen PDF viewer, real-time citation validation badges, and live CSCS inference telemetry.
📺 Demo Video: Watch the 3-minute Walkthrough
🔗 GitHub: https://github.com/vero-code/severin
🛠️ How I Built It
- Supercomputing Brain: Powered by
swiss-ai/Apertus-v1.5-8Brunning remotely on NVIDIA Grace Hopper GH200 nodes at CSCS Alps in Lugano via an OpenAI-compatible API (temperature=0.0). - Data Model & Guardrails: Built with Pydantic v2 using strict schemas (
extra="forbid"). I engineered a tailored Defensive Normalizer to handle multilingual JSON formatting quirks from smaller LLMs without masking live errors. - Multi-Modal Document Parsing: Implemented physical page extraction via
pypdf. For pure raster scans with zero extractable text (e.g. Canton Obwalden), I designed the system to automatically trigger Apertus Vision OCR viaPillowimage extraction. - Backend & Frontend: Developed with FastAPI (OpenAPI 3.1) and a sleek React 19 single-page application.
- Turnkey Reproducibility: Packaged the entire local stack into a multi-stage Docker container with a single entrypoint:
make run.
🧗 Challenges I Ran Into
- Multilingual Structural Idiosyncrasies: Handling German interpellation numbering, French formal questions, and Italian parliamentary addresses required careful prompt engineering to ensure Apertus 8B preserved exact sentence structures without truncating text.
- Scientific Integrity vs. Deceptive Fallbacks: I committed to strict academic honesty: live CSCS inference is never deceptively masked by cached mocks. If an API call fails or hits limits, it is logged openly.
- Physical Page Anchoring: Matching model text snippets to physical PDF bounding boxes required me to build custom whitespace and ligature normalization to ensure 100% character-level provenance against source documents.
🏆 Accomplishments I'm Proud Of
- Built solo in 7 days: A complete, production-grade proof-of-concept from scratch.
- 100.0% Verifiable Provenance: Achieved 19/19 verbatim citation matches across a 5-jurisdiction multilingual benchmark (CHE, ZH, GE, TI, GR).
- Extreme Efficiency on CSCS Alps: Cold latency averages 5.12s, dropping to 3.16s on warm cache requests thanks to vLLM PagedAttention prefix caching, at an average cost of only 0.000047 CHF per document.
- Complete Academic Report: Authored a rigorous 6-page technical report following the competition standard.
🎓 What I Learned
- How to architect reliable, deterministic workflows with the CSCS Alps supercomputing infrastructure and take advantage of PagedAttention token caching.
- That small, sovereign 8B open-weight models, when paired with strict Pydantic schemas and algorithmic verification, can achieve zero-hallucination document extraction rivaling proprietary 70B+ models at a fraction of the cost.
Log in or sign up for Devpost to join the conversation.