Bünzli.Space

A sovereign, open-source AI assistant for Zürich. Swiss models, Swiss servers, Swiss law, live local data, privacy you can verify from your browser's network tab.

Live at buenzli.space — landing page + closed beta. Chat lives at /chat/.

Built solo by a non-engineer PM, with Claude as a pair programmer. Disclosure rather than disclaimer: the tooling is American, the product is Swiss. Inference, storage, jurisdiction, and every byte of user data stay inside .ch.

Status — April 2026

Recent shipped work: see CHANGELOG.md.

Why

Four constraints shape every technical decision:

1. Affordable. Free during beta; modest subscription at worst long-term. Never ad-supported.
2. Locally competent. Trained for Swiss German, grounded in live Zürich APIs — not Switzerland-in-general.
3. Privacy-by-default. Swiss jurisdiction end-to-end, encrypted client-side, anonymised, no training on user data.
4. Open. MIT-licensed code, auditable infrastructure, zero telemetry.

For every layer there was a cheaper US-hosted alternative. They were rejected on principle.

Architecture

Bünzli runs across two physical machines in Infomaniak's Zürich datacentre, organised as pods — one Docker Compose project per coherent slice of functionality.

   ┌───────────────────────────────────────────┐         ┌─────────────────────────────┐
   │  Infomaniak VPS  (buenzli.space)          │         │  Infomaniak GPU instance    │
   │                                           │         │  (gpu.buenzli.space, A2)    │
   │  ┌─────────────────────────────────────┐  │  HTTPS  │                             │
   │  │  bunzli-app pod                     │──┼─────────┼─▶ bunzli-gpu pod            │
   │  │   • bunzli-app-edge   (Caddy/TLS)   │  │         │   • bunzli-gpu-edge         │
   │  │   • bunzli-app-api    (FastAPI)     │  │         │   • bunzli-gpu-embed        │
   │  │   • bunzli-app-landing (Astro)      │  │         │   • bunzli-gpu-rerank       │
   │  └─────────────────────────────────────┘  │         │   • bunzli-gpu-qdrant       │
   │                                           │         │   • bunzli-gpu-llm (staging)│
   │  ┌─────────────────────────────────────┐  │         │                             │
   │  │  bunzli-app-iam pod (Zitadel SSO)   │  │         └─────────────────────────────┘
   │  └─────────────────────────────────────┘  │
   └───────────────────────────────────────────┘

Full operational write-up: docs/infrastructure.md.

Pods

Stack

No piece of the runtime stack is physically outside Switzerland.

Privacy, concretely

The promise "your data stays safe" is cheap. Here's what it actually means on Bünzli:

• Client-side encryption. Chat history is AES-256-encrypted in the browser before it touches the network. The server persists ciphertext only.
• Pseudonymous identity. Sessions are tied to an opaque Zitadel user ID, not an email or IP in request logs.
• No cross-border egress. Every request terminates in .ch. Verify it yourself — DevTools → Network tab while chatting.
• No analytics, no telemetry, no tracking pixels. The only per-session log entry is a token count for billing reconciliation.
• No training on user data. Ever.
• Open source. The claim is only as good as the code backing it. Read it. Break it. Open an issue.

Security posture

• Network — Infomaniak anti-DDoS, OpenStack security groups (22/80/443 only), key-only SSH, no root login.
• Application — JWT bearer auth on /v1/chat/completions, ModSecurity + OWASP CRS planned, 10 req/s rate limit on backend routes.
• Auth — Zitadel brute-force protection, OIDC + PKCE from the SPA.
• Monitoring — weekly OWASP ZAP baseline scan via GitHub Action.
• No CDN. Cloudflare et al. were rejected to keep the Swiss-jurisdiction story clean.

Connectors

Each connector is a small Python module in backend/connectors/ mapping a public Swiss/Zürich data source into a tool the agent can call.

Missing a source? Open an issue or PR.

Knowledge base

A from-scratch rebuild of Bünzli's RAG corpus is in flight. Phase 1 (ingest, normalise, chunk) finished April 2026 with 41,714 chunks across a 10-category taxonomy — federal/cantonal law, city services, transport, history, culture, civic life. Phase 2 (embedding + retrieval acceptance on the GPU pod) is the current sprint. Spec lives in docs/knowledge_base.md.

History sources are deliberately curated for Bünzli's voice (turns of phrase, local references), not as a Q&A reference — the model is meant to feel like it grew up here, not parrot Wikipedia.

Roadmap

Near term (next ~6 weeks):

• Retrieval online. Embed Phase 1 chunks, stand up Qdrant + reranker on bunzli-gpu, pass acceptance evals.
• Orchestrator v1. Code-orchestrated chitchat / RAG / tool-call routing — Apertus voice + Mistral tool-call, browser-local state. See docs/orchestration_architecture.md.
• Self-hosted auth live. Zitadel pod surfaced at auth.buenzli.space, wired into the SPA.

Medium term:

• Action-taking agents. Three first-party tools beyond Q&A: restaurant booking (Quandoo), waste-pickup scheduling (ERZ), street-defect filing (Züri-wie-neu).
• Multi-language output. Full CH-DE / DE / EN / FR / IT via Supertext API.
• Public beta. Summer 2026, opening with 50 seats.

Longer term:

• Swiss-German fine-tune once we have enough labelled in-domain data. Until then: glossary in the system prompt, no runtime translation.
• Native mobile if the web app finds product-market fit.

Repository layout

zuribot/
├── api_server.py              # FastAPI entry point (OpenAI-compatible)
├── backend/
│   ├── agent.py               # Plain-Python orchestrator
│   ├── tools/                 # Tool definitions + dispatch
│   ├── connectors/            # Zürich data source adapters
│   └── models/                # Pydantic schemas
├── frontend/
│   ├── landing/               # Astro site (landing + /chat/ shell)
│   └── chat/                  # React + assistant-ui chat island
├── deploy/
│   └── pods/
│       ├── app/               # bunzli-app — edge + api + landing
│       ├── app-iam/           # bunzli-app-iam — Zitadel
│       └── gpu/               # bunzli-gpu — embed + rerank + qdrant + staging LLM
├── kb/                        # Knowledge base ingest pipeline + Phase 1 output
├── docs/                      # Infrastructure, KB, orchestration, staging spec
├── evals/                     # Eval suites
└── scripts/                   # One-shot tooling (embed, audits, …)

Running locally

Prerequisites

• Python 3.11+, Node.js 20+, Docker
• Infomaniak AI Tools API key (or any OpenAI-compatible endpoint)

Backend

python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt          # core
pip install -r requirements-heavy.txt    # torch, sentence-transformers, chromadb (only if running KB locally)
cp .env.example .env                     # fill secrets locally; never commit
python3 api_server.py

API listens on :8000 (OpenAI-compatible /v1/chat/completions).

Frontend

cd frontend/landing && npm install && npm run dev

Astro serves on :4321. The /chat/ route proxies to the API.

Full pod stack (TLS + reverse proxy)

docker compose -f deploy/pods/app/docker-compose.yml up -d --build

Caddy can't issue real Let's Encrypt certs locally — that's expected; hit http://localhost:80.

Deployment

Production pods run on an Infomaniak VPS in Zürich, managed via Docker Compose. The bunzli-app pod is the public-facing one.

# Ship .env if it changed:
scp .env bunzli@<vps>:/home/bunzli/zueribot/.env

# Pull + rebuild:
ssh bunzli@<vps> 'cd ~/zueribot && git pull && cd deploy/pods/app && docker compose up -d --build'

# Smoke-test:
curl -sS -o /dev/null -w "site: %{http_code}\n" https://buenzli.space/
curl -sS -o /dev/null -w "api:  %{http_code}\n" https://buenzli.space/zuribot/health

Secrets live in the repo-root .env, never committed. See .env.example for required keys, docs/infrastructure.md for the full operational picture, and per-pod READMEs under deploy/pods/ for pod-specific deploy notes.

Pre-deploy rule: any change to a pod's compose / Caddyfile / Dockerfile / env layout must be brought up on Docker Desktop locally first. Reason: we shipped the pod restructure straight to prod once and took the site down for five minutes.

Cost model

• Free tier — casual use stays free during and after beta.
• Subscription tier — CHF-scale monthly fee for heavy users, only if someone opts in.
• No ads. Ever.

A 70B open-weight model served per-token via Infomaniak is cheap enough that current beta usage sits comfortably inside a personal budget. Serving cost analysis: docs/token_cost_strategy.md.

On using AI to build it

Most of Bünzli was written with Claude as a pair programmer. The tooling is Anthropic's; the product is Swiss. The inference, storage, jurisdiction, and live user data — the things that matter for the privacy claim — are Swiss-hosted end to end.

Contributing

PRs welcome, especially:

• New connectors for public Zürich / Swiss data sources
• Swiss German / Züridütsch improvements
• Prompt + eval additions
• Security reviews

Please don't open PRs that add tracking, analytics, or cross-border data flows — they conflict with the design goals and will be closed.

Contact

• Product: buenzli.space
• Email: [email protected]

License

MIT. See LICENSE.