AddrLens

Should I rent it? Know your Berlin neighbourhood in thirty seconds — schools, transit, noise, air, history — before signing your next lease.

Live at addrlens.de · MIT-licensed · No cookies · No signup · GDPR-clean

---

What it is

AddrLens takes any Berlin address and, in under 30 seconds, cross-checks it against 20+ open datasets published by the Berlin Senate, VBB, BVG, and OpenStreetMap:

• Assigned Grundschule (Einschulungsbereich)
• Kitas within walking distance
• S-Bahn / U-Bahn / Tram / Bus reachability (VBB + OSM)
• Umweltatlas noise, air (NO₂), heat
• Green refuge, drinking fountains, public pools
• Fire station coverage, quiet zones, Milieuschutzgebiete
• Neighbourhood profile (GESIx 2022 socioeconomic band)
• Nearby historic markers / Stolpersteine
• German admin office reach (Bürgeramt, Finanzamt, Standesamt, LEA, Arbeitsagentur)

Four curated Life Mode lenses (Young Family, Newcomer, Quiet Living, Commuter) surface the tiles most relevant to each situation. A Raw view lets anyone browse every dataset with full source attribution and licence tags.

Each lens ships an AI Insight panel — the model turns per-lens deterministic facts into a plain-English paragraph. Prod calls Cloudflare Workers AI (@cf/meta/llama-3.1-8b-instruct-fast); local dev falls back to mlx-lm on Apple Silicon or llama-cpp-python on Linux. Address-agnostic prompts + a 7-day server-side cache mean neighbours share the same LLM-derived text — cutting external calls and preserving privacy.

Screenshots

Life Lens: Young Family	AI Insight

Raw view (all datasets, full attribution)	Life Lens compare (tier chips)

Data sources & licences

Source	Datasets used	Licence
Berlin Geoportal / FIS-Broker (gdi.berlin.de)	Schulen, Kitas, Krankenhäuser, Feuerwehr, Milieuschutz, Umweltatlas, Baumbestand, Trinkbrunnen, Grünanlagen, Bezirksgrenzen	dl-de/by-2-0 · dl-de/zero-2-0
VBB Verkehrsverbund Berlin-Brandenburg	S-Bahn / U-Bahn / Tram station coordinates	CC-BY-4.0
BVG Berliner Verkehrsbetriebe	Bus stops, service data	dl-de/by-2-0
Geofabrik weekly extract of OpenStreetMap Berlin	International food, coworking, English clinics, language schools, libraries, Packstationen, Wochenmärkte, historic markers, address autocomplete	ODbL 1.0, © OpenStreetMap contributors
Berlin Senate GESIx 2022 (Gesundheits- und Sozialstrukturatlas)	Neighbourhood socioeconomic band per Planungsraum	dl-de/zero-2-0

Every card in the product surfaces its dataset name and licence tag. The full attribution list also lives in legal/impressum.md.

Stack

• Backend: Python 3.11+, FastAPI, Shapely, httpx, slowapi for rate limiting, cachetools for the history TTL cache
• Inference: Cloudflare Workers AI (@cf/meta/llama-3.1-8b-instruct-fast) in prod; dev falls back to mlx-lm on Apple Silicon or llama-cpp-python (Qwen 2.5 1.5B Q4_K_M GGUF) on Linux — selected by INFERENCE_BACKEND
• Frontend: vanilla HTML / CSS / JS, Leaflet for maps. No bundler, no framework, no build step.
• Analytics: self-hosted Umami (cookieless, no consent banner needed)
• Errors: Sentry EU region, send_default_pii=False, address + IP redacted in before_send
• Deploy: Docker Compose on a Hetzner CX22 (Debian 13, Falkenstein DE) behind a Cloudflare Tunnel; weekly OSM refresh timer, per-endpoint WAF + rate-limit rules

Repository layout

The active tree is v0.1/. Historical prototypes (phase0/–phase3/) are archived under docs/history/prototypes/ for lineage — many v0.1/app/core/* modules are ports of phase3/server.py and cite it in their comments.

Directory	Purpose
v0.1/app/	FastAPI app — one process per city, selects CityConfig at boot via CITY=<slug>
v0.1/inference/	Shared LLM service (city-agnostic). Prod: Cloudflare Workers AI (INFERENCE_LOCAL_BACKEND=off). Dev: mlx (Apple Silicon) or llama.cpp (Linux) via INFERENCE_BACKEND
v0.1/web/	Single-page frontend (no bundler, no build step) — index.html + legal pages + static/app.js (126-LOC ES-module bootstrap) + 20 modules under static/modules/ (api, state, dom, maps, lens/, panels/)
v0.1/scripts/	Weekly OSM refresh (extracts amenities + addresses from Geofabrik)
v0.1/ops/	Dockerfiles, Cloudflare Tunnel config, systemd units, deploy scripts
v0.1/docs/	Deploy playbook, WAF setup, social-embed playbook
legal/	Impressum + Datenschutzerklärung (bilingual DE/EN, DDG §5 + DSGVO Art. 13 compliant)

Local development

Two paths — pick one.

1. Bare uvicorn on macOS Apple Silicon (fastest for iteration; uses Metal GPU via mlx-lm)

cd v0.1

# One-time on a fresh clone — same venv as the pytest suite (see ## Tests)
python3.13 -m venv .venv
uv pip install --python .venv/bin/python -e ".[dev]"
uv pip install --python .venv/bin/python -e "./inference[dev]"      # mlx-lm

# Terminal 1 — inference on :8080
INFERENCE_BACKEND=mlx .venv/bin/uvicorn inference.main:app --port 8080

# Terminal 2 — app on :8003
CITY=berlin INFERENCE_URL=http://localhost:8080 \
  .venv/bin/uvicorn app.main:app --port 8003

Open http://localhost:8003/.

2. Docker Compose (matches prod)

cd v0.1

# Fetch the GGUF language model once (~940 MB)
mkdir -p models
curl -L -o models/Qwen2.5-1.5B-Instruct-Q4_K_M.gguf \
  "https://huggingface.co/bartowski/Qwen2.5-1.5B-Instruct-GGUF/resolve/main/Qwen2.5-1.5B-Instruct-Q4_K_M.gguf?download=true"

# The dev override mounts models/ + data/osm into the containers
docker compose up -d --build

Open http://localhost:8001/.

Neither path needs a Cloudflare Workers AI token — the app happily runs against the local llama.cpp fallback for every LLM call.

Tests

Two complementary test tiers:

pytest suite — 211 tests under v0.1/tests/ (unit + contract + opt-in integration). Runs in ~1 s. CI green on every push and PR via .github/workflows/test.yml.

cd v0.1

# One-time on a fresh clone
python3.13 -m venv .venv
uv pip install --python .venv/bin/python -e ".[dev]"

# Every time
.venv/bin/pytest                            # unit + contract, no network
.venv/bin/pytest -m integration             # opt-in live Berlin Geoportal
.venv/bin/pytest --cov=app                  # coverage report

Selfchecks — every module ships a pure __main__ block that asserts its own contract; a single app.selfcheck orchestrator runs everything plus live-network checks against gdi.berlin.de. Kept alongside pytest for sub-second dev-loop feedback on a single module.

cd v0.1

# Pure asserts — no network, no model load, sub-second each
python -m app.core.scorer            # tier boundaries + lens composers
python -m app.core.rate_limit         # CF-Connecting-IP key extraction
python -m app.core.address_index      # umlaut + street-suffix normalisation
python -m app.routes.lens_insight     # per-lens AI template registration

# Live-network — pulls WFS layers, needs gdi.berlin.de reachable
python -m app.selfcheck

Deployment

See v0.1/docs/hetzner-deploy.md for the full playbook — Debian 13 host bootstrap, Cloudflare Tunnel, secrets, systemd OSM refresh timer, Sentry + Umami wire-up.

Related runbooks:

• v0.1/docs/social-embed-og-image.md — social preview cards
• v0.1/docs/cloudflare-waf-setup.md — free-tier WAF rules

Roadmap

Concrete deliverables for the next six months, in rough order:

• Snapshot fallback for gdi.berlin.de. The Berlin Senate WFS had multi-day maintenance in August 2026; snapshot-based resilience is a prerequisite for treating this as public infrastructure.
• Second city (Hamburg, then München). The architecture already picks the city per CityConfig at boot; each new city needs a data-source map and glossary translation.
• Public JSON API with a documented schema, so other Berlin civic-tech projects can reuse the aggregated address data without re-implementing 20 WFS clients.
• Mietspiegel Wohnlagen integration — the only openly licensed rent-level signal per address.
• WCAG 2.2 AA accessibility audit — screen-reader coverage, keyboard navigation, high-contrast mode.
• Historical time-series (e.g. "has noise on this address improved over the last 5 years?") once the data horizon allows.

Contributing

Bug reports, data-source additions, and small documentation fixes are welcome. See CONTRIBUTING.md for the layout, code style, and selfcheck workflow.

Licence

MIT — see LICENSE. Data sources retain their own licences (see the table above).

Acknowledgements

• Berlin Senate departments for publishing the datasets that make this possible under open licences.
• ODIS Berlin / Technologiestiftung Berlin for maintaining the ecosystem around Berlin open data.
• OpenStreetMap contributors — every historic marker, café, and bus stop.
• VBB Verkehrsverbund Berlin-Brandenburg and BVG for open-data-licensing the transit network.

Contact

• Product: https://addrlens.de/
• Feedback: [email protected]
• Issues + feature requests: GitHub Issues