SiteScout

Inspiration

Opening a business is a major financial commitment, yet where to open usually comes down to gut feel, a few hours of research, or an expensive consultant. A busy high street can mean brutal competition, while a quieter neighbourhood may have more customers, lower costs and nobody serving them. We wanted a first-time founder in Cork to see the evidence before signing a lease.

What it does

You pick what you're opening (café, barber, gym, restaurant or takeaway, convenience store), who your customers are, and your budget. SiteScout scores about 589 areas of Cork, each around 350 m across, on four things: demand, competition, access and cost. It then shows:

a heat map of every area, which you can colour by any of the four scores
the top five locations, kept at least 600 m apart so they aren't five neighbours
for each pick, an evidence table against the Cork median (residents, offices, bus stops, competitors, distance from the centre) and the names of the nearest competitors
a side-by-side comparison of all five picks
a budget switch that re-ranks the picks live
real commercial listings near each pick, from a snapshot of 50 Cork properties on Daft.ie
a team of AI agents that explains each pick in plain English, with the risks

How we built it

The ranking is plain code, not AI. The same inputs always give the same order, and every number can be checked. Each area is percentile-ranked against every other Cork area. Demand is weighted at 40%, competition 25%, access 20%, and cost 30%, 18% or 8% depending on budget.

AI only explains. Four Gemini specialist agents (demand, competition, access, cost) read their slice of the evidence in parallel. A challenger agent argues against their claims, and a reporter writes the verdicts. Every verdict is then checked in code against the numbers in the data, and a verdict that doesn't match is replaced with a data-only one. With no API key, the app falls back to data-only explanations.

Data: CSO Census 2022 small areas for residents by age band, OpenStreetMap via the Overpass API for competitors, offices, bus stops, shops, universities and tourist sites, and a Daft.ie snapshot for listings. We pulled it once and cached it locally, so the app runs offline and gives identical results every time.

Stack: Python and FastAPI for the backend, React, Vite and Leaflet for the frontend, and Google Gemini for the agents. We split the work three ways around a shared JSON API contract, so scoring, agents and frontend could be built in parallel.

Challenges

Cost data doesn't exist openly. We found no open commercial-rent dataset for Cork, so the cost score is distance from the city centre, and we label it as a proxy everywhere in the app.
Keeping AI honest. Agents can invent figures. We confine them to supplied numbers, and a code check compares each verdict against the data. Early on it wrongly rejected every verdict because it read grid-cell IDs like r9c29 as numbers, which we fixed.
Live model choice. Our first Gemini model wasn't available to new keys, and the next one took around 50 seconds per call, too slow for a demo. A lighter model brought the full agent run down to about 7 seconds.

Limits we state openly

SiteScout gives a better-informed starting point, not a forecast.

Cost is a proxy, not real rent.
OpenStreetMap is community-mapped, so counts can be incomplete.
Scores are relative: they rank Cork areas against each other and don't predict revenue or footfall.
Universities are single points, so large campuses are under-represented in the student signal.
The number check catches invented large figures (rents, footfall, prices) but not every small number.
We haven't checked lease availability, planning use or local charges.

What we learned

Keeping scoring deterministic and using AI only for explanation made the product easier to trust, test and demo. A shared API contract let three people build in parallel without stepping on each other.

What's next

Show it to five Cork founders and ask what they would have paid to see this before signing their lease. The census and map data cover all of Ireland, so each new city is a data pull, not a rebuild. The missing piece for a real business is commercial rent and lease data, which isn't open, so we'd look at a partnership with a property-data provider.