Ask where to open your business. Get a map, not an essay.
WhereHouse is a geospatial chat agent for site selection. You ask, in plain words, "Where should I open a bakery in Berlin?" — and the answer assembles itself on a live map: the competition drops in as dots, a demand-vs-supply opportunity surface fades in, the three best sites pin themselves, a real 10-minute walking catchment blooms around the winner, and a momentum sparkline says whether the market is rising or saturating. The agent writes at most two sentences. Everything else is drawn.
Built for the ClickHouse × Trigger.dev Virtual Summer Hackathon 2026 — theme "Beyond the Wall of Text", judged on the ratio of insight to words.
Live demo: https://app.slim-shaggy.com — and the app is served out of ClickHouse itself (see ADR-003).
Demo cities: Berlin · Amsterdam · Belgrade.
Every frame is the real deployed app at app.slim-shaggy.com assembling a live answer — no mockups, no stock footage. One question per city, then the follow-ups that make it a consultant, not a calculator: ask for the biggest markets, rule out the saturated turf, or zoom into one neighbourhood, and the pins re-rank themselves. Click a still to watch on YouTube (~2 min each):
Prefer the raw files? All three are attached to the demo release.
"Where should I open this?" is an inherently spatial question. Answering it with a paragraph is the exact failure "Beyond the Wall of Text" is about. So the product has a hard rule, enforced in the agent's own prompt: the map is the answer; prose is a defect. The model never sees the geometry — it gets row counts and small summaries — so it cannot recite the data back at you even if it wanted to. Its two-sentence caption is only allowed to say what the map cannot show: the pattern behind the answer.
Not a cache, not a sidecar. Every layer on the map is a GeoJSON string assembled in SQL and streamed straight to the browser.
- 75M+ Overture Maps POIs, queried in place from public S3 with
s3()— no ingestion pipeline. 211,818 POIs for the three cities are materialised intogeo.placeswith H3 indexes computed inMATERIALIZEDcolumns. (ADR-002) - H3 hexagonal scoring. Demand (475,535 Kontur population cells) joins supply (POIs in an
h3kRing) and accessibility (Valhalla isochrones snapped to H3) on cell-id equality — no interpolation. The GAP score is one explainable SQL formula, derived per city and category at query time (p95-normalised, no magic constants). - The app is served from a
MergeTreerow. The static bundle lives inweb.assets; a thin Cloudflare Worker streams it and the browser then queries ClickHouse directly (CORS is open) for the map's data. GeoJSON is built by hand withtoJSONString/arrayStringConcatbecause Cloud 26.4 predates the nativeGeoJSONformat. (ADR-003) - Real ClickHouse feature depth, where it earns its place: a
Dictionary+dictGetFloat32UDF for complementary-business affinity; an incrementalAggregatingMergeTreematerialized view (category_momentum) rolling OSM edit-history into per-cell monthly momentum;ReplacingMergeTree+FINALfor the saved-site store. - Your own state, joined against the world: a site you saved seconds ago is re-scored against today's 75M-POI market on H3 equality. It lived in a ClickHouse-managed Postgres replicated back by ClickPipes CDC for the hackathon's OLTP+OLAP prize (ADR-004); it now lives in ClickHouse itself, and the join is unchanged (ADR-005).
- The agent exposes 14 tools. Eight build the answer (
findCompetitors,scoreArea,showBuiltCapacity,rankSites,showCatchment,categoryTrend,saveSite,compareSavedSites); six let the agent operate the UI itself (toggle layers, move the re-weight sliders, focus a pick, highlight the best/worst cell, rewind to a past run, export a PDF). rankSitesis a consultant, not a fixed top-3. Beyond the balanced best three it ranks under a strategy lens (biggest market / least competition / best foot traffic), shows where not to open (the most saturated cells), ranks inside a named district, and pages up to six pins — all over the same scored cells, with the balanced default unchanged.- The progressive map is the innovation. Each tool emits a
data-mappart with a stable id; re-writing that id updates the part in place instead of appending. That is what makes the map fill in wave by wave while the agent is still working, instead of appearing all at once at the end. (ADR-001) - A hard ~1 MiB per-record stream cap is designed around, not hit. An over-budget layer
(e.g. all 6,664 food-and-drink points) never touches the stream: it lands in a
web.layersrow and the part carries a handle, which the browser resolves directly from ClickHouse. - Deployed to Trigger.dev's managed prod environment, so the live site works with no local
process running (
infra/deploy-trigger.sh).
Trigger.dev chat.agent()
you ask ──▶ ┌──────────────────────────────────────────┐
│ findCompetitors → scoreArea → rankSites │
│ → showCatchment → categoryTrend │
└───────────────┬──────────────────────────┘
│ (model sees only row counts + summaries)
data-map / data-trend parts, stable ids, updated in place
│
▼
ClickHouse Cloud ◀── GeoJSON built in SQL ──▶ MapLibre (the answer assembles)
• geo.places (Overture POIs, H3)
• geo.population (Kontur demand)
• geo.isochrones (Valhalla walk catchments, precomputed)
• geo.districts (Overture divisions → real place names)
• geo.affinity (Dictionary UDF)
• geo.category_momentum (incremental MV)
• app.saved_sites (your saved sites — the same DB, joined on H3)
The first screen is a Claude/ChatGPT-style omnibar; on your first question it transitions into the map dashboard, where the layers draw themselves.
| Layer | Source | Volume |
|---|---|---|
| Places, districts | Overture Maps via public S3 — Places CDLA Permissive 2.0, Divisions ODbL | 211,818 POIs · 4,075 district shapes |
| Population (demand) | Kontur (CC BY 4.0) | 475,535 H3 res-8 cells |
| Walk catchments | Valhalla on OSM (ODbL), precomputed | 63,479 isochrones · 20,724 origins |
| Momentum | ohsome/OSM edit history (ODbL) | monthly series → incremental MV |
| Basemap | Protomaps PMTiles / OpenStreetMap | 3 cities < 100 MB |
Every number the agent utters is one a tool returned; place names come from real point-in-polygon against Overture divisions, never invented. Editorial signals (the affinity weights) are labelled as editorial in every code path and in the UI — never presented as measurement.
cp .env.example .env # fill in ClickHouse + Trigger.dev credentials
./infra/check-env.sh # verifies every credential against the live services (~5s)
cd web
pnpm install
pnpm dev # the app on http://localhost:3000
pnpm exec trigger dev # the chat.agent() task, in a second terminalProvision the cloud from nothing (idempotent, all via the Cloud REST API):
./infra/provision.sh # ClickHouse service + the app schema
./infra/deploy-app.sh # build the static bundle → load into ClickHouse → deploy the Worker
./infra/deploy-trigger.sh # ship the agent to Trigger.dev prod and wire the Worker to it| Path | What |
|---|---|
web/ |
Next.js app (static export). src/trigger/ = the agent + SQL; src/components/chat.tsx = the whole UI |
infra/ |
Everything provisioned as code — ClickHouse, basemap, Valhalla, deploys |
db/ |
ClickHouse schemas and loaders |
video/ |
The demo video pipeline — scenario, voiceover script, Remotion + Playwright |
docs/PLAN.md |
Day-by-day plan, what's cut, risk register |
docs/architecture/ |
ADR-001…005 — read before changing direction |
.specify/memory/constitution.md |
Project constitution — verify against the live system, prove the riskiest path first |
- The map is the response — layers stream in place through
chat.agent(), wave by wave. - ClickHouse is the web server — the app is a database row; the browser talks to the DB.
- Your state × the world in one join — a site saved seconds ago, scored against 75M POIs.
- Honesty as a feature — absent data is shown as absent; unmeasured cells are labelled, not zero-filled; the agent is structurally unable to invent a place name.
MIT — see LICENSE. Map data © OpenStreetMap contributors via Overture Maps and Protomaps; population © Kontur (CC BY 4.0).