GenAI Zürich Hackathon 2026 · Qdrant Challenge
Compliance teams at Swiss private banks spend weeks manually verifying that their internal procedures align with FINMA circulars.
"A compliance officer spends an average of 3 hours per week manually searching through FINMA texts."
RegBridge automates this analysis using generative AI and Qdrant vector search.
RegBridge is a conversational AI assistant that enables a compliance officer to:
- 📄 Query FINMA regulations in natural language (circulars 2016/7, 2023/1, 2025/2)
- 📁 Upload internal documents (policies, procedures, contracts) → indexed in real time
- 🔀 Cross-reference official regulations + internal documents in a single query
- ⚡ Get sourced answers with precise article and page references
- 📊 Export cross-analysis results as a PDF compliance report
- 🌍 Multilingual : FR · EN · DE · IT
This is the technical heart of RegBridge. Two dedicated Qdrant collections are queried simultaneously to detect compliance gaps between official regulations and internal procedures.
┌─────────────────────────────────────────────────────────────────┐
│ QDRANT CLOUD — Dual Collection │
│ │
│ ┌──────────────────────────┐ ┌──────────────────────────────┐ │
│ │ Collection 1 │ │ Collection 2 │ │
│ │ finma_docs │ │ internal_docs │ │
│ │ │ │ │ │
│ │ • Circ. FINMA 2023/1 │ │ • Internal procedures │ │
│ │ • Circ. FINMA 2016/7 │ │ • Policies & regulations │ │
│ │ • OBA-FINMA │ │ • Uploaded PDFs │ │
│ │ • CDB 2020 │ │ • Auto-chunked + indexed │ │
│ │ │ │ │ │
│ │ HuggingFace Embeddings │ │ HuggingFace Embeddings │ │
│ │ 384 dimensions │ │ 384 dimensions │ │
│ └──────────┬───────────────┘ └──────────────┬───────────────┘ │
│ │ │ │
│ └──────────────┬───────────────────┘ │
│ │ │
│ Cross-Query Engine │
│ (LangChain MergerRetriever) │
│ → Gap detection · Sourced citations · FR/DE/EN/IT │
└────────────────────────────┬────────────────────────────────────┘
│
┌──────────────▼──────────────┐
│ Groq API · LLaMA 3.1 8B │
│ (response generation) │
└──────────────┬──────────────┘
│
┌──────────────▼──────────────┐
│ FastAPI Backend · Python │
│ POST /query │
│ POST /cross-query │
│ POST /upload │
└──────────────┬──────────────┘
│ HTTPS · ngrok
┌──────────────▼──────────────┐
│ React Frontend · Lovable │
│ regbridge.lovable.app │
└─────────────────────────────┘
| Single-collection RAG | RegBridge dual-collection |
|---|---|
| Can only answer questions about one corpus | Queries FINMA + internal docs simultaneously |
| No gap detection | Detects misalignments between regulation and practice |
| Generic compliance Q&A | Actionable recommendations per institution |
| Replicable by any LLM | Private document indexing = institutional value |
React interface deployed via Lovable
FastAPI backend on Google Colab + ngrok
⚠️ To test: run notebook Step 2 first to activate the backend
| Tab | What it does |
|---|---|
| 💬 FINMA Question | Natural language query on indexed FINMA circulars · sourced answers with PDF + page |
| 🔀 Cross-analysis | Simultaneous query of both Qdrant collections · compliance gap detection · export results as PDF report |
| 📁 Internal documents | Upload PDFs · real-time Qdrant indexing · library with audit trail |
| Layer | Technology | Role |
|---|---|---|
| Frontend | React · TypeScript · Lovable | User interface (FR/DE/EN/IT) |
| Backend | FastAPI · Python · Uvicorn | REST API |
| Embeddings | HuggingFace paraphrase-multilingual-MiniLM-L12-v2 |
Multilingual vectorization (384 dim) |
| Vector DB | Qdrant Cloud | Dual-collection vector storage + search |
| RAG | LangChain · MergerRetriever | Cross-collection retrieval pipeline |
| LLM | Groq API · llama-3.1-8b-instant |
Response generation (<1s latency) |
| Backend deploy | Google Colab · ngrok | Public FastAPI server |
| Frontend deploy | Lovable | Web interface |
| Collection | Content | Embedding model | Usage |
|---|---|---|---|
finma_docs |
FINMA Circulars 2016/7, 2023/1, OBA-FINMA, CDB 2020 | paraphrase-multilingual-MiniLM-L12-v2 |
Regulatory questions |
internal_docs |
Internal documents uploaded by the user | paraphrase-multilingual-MiniLM-L12-v2 |
Cross-reference queries |
QDRANT_URL=https://xxx.qdrant.io
QDRANT_API_KEY=your_qdrant_key
GROQ_API_KEY=your_groq_key
NGROK_AUTH_TOKEN=your_ngrok_token
VITE_API_URL=https://xxx.ngrok-free.app# Step 1 — Index FINMA documents (run once)
# Open notebooks/Step 1 → Run all cells
# FINMA circulars indexed in Qdrant finma_docs ✅
# Step 2 — Start the server (each session)
# Open notebooks/Step 2 → Run all cells
# Copy the public ngrok URL → paste into the frontendNo installation required — live at regbridge.lovable.app
| # | Notebook | Description |
|---|---|---|
| 1 | Step 1 — FINMA PDF Ingestion | Load, split and index FINMA circulars into finma_docs |
| 2 | Step 2 — Full Server | FastAPI + dual-collection RAG + ngrok + internal document upload |
| Version | Features | Status |
|---|---|---|
| v1.0 MVP | FINMA RAG · Internal docs upload · Dual-collection cross-analysis · PDF report export · 4 languages | ✅ Today |
| v2.0 | DORA · DPA/GDPR · Basel III/IV · Multi-user with roles · Cross-session memory | 🔒 Post-hackathon |
| v3.0 | 40+ regulatory frameworks · Auto-alerts · Core banking integration | 🔒 Long-term |
| Name | Role |
|---|---|
| @dagracarui25-hash | Senior Systems Engineer · Private Bank Geneva |
MIT — Built for GenAI Zürich Hackathon 2026 · Qdrant Challenge