Skip to content

Live · private preview

Trestate

Location intelligence for anyone about to sign for a shop. Type an address and a trade: Trestate grades the parade around that door, maps the competition within a five- and a ten-minute walk, and in the Paris region shows what traded there before.

Role
Founder · Location Intelligence Engineer
Period
Aug 2026 – present
Stack
Go · PostgreSQL / PostGIS · H3 · Python + DuckDB · and 9 more
Place
London & the Paris region
Ask at 95 Rue de Charenton: A+ for a café, the cafés within a five- and a ten-minute walk, and the nearest station.

Ask at 95 Rue de Charenton: A+ for a café, the cafés within a five- and a ten-minute walk, and the nearest station.

44,885parades graded

A buyer brings an address and an asking price, and leaves knowing what the location supports. Someone opening their first shop learns how the street they have in mind ranks for their trade before they sign the lease. A chain brings a format and leaves with the next sites.

01 / 06

Who it is for

Agents and brokers attach a Dossier to a listing in the Paris region, with London's on the way.

Lenders and accountants read the location behind the file on their desk: the grade, the competition, the station, and in the Paris region the address's record and what businesses sold for nearby.

From trestate.pages.dev: lending or advising, a file on the desk and an address in it.
Who it is for
Local sites ranked for café, bakery, pharmacy and beyond. London and Paris.
02 / 06

The product

One address, one trade, and the street around it, read.

Ask takes the question in plain English or French, misspelt street types and abbreviations included, and resolves it to a named door on a named street. Gemini narrates the answer in a sentence; the grade, the map and the counts come from the data underneath it.

  • Grade the parade. The parade (the shops around a door) is graded for your trade, against all the parades Trestate grades for that trade.
  • Who you'd be up against. The other shops of your trade on record within a five- and a ten-minute walk land on the map, with the chains among them named. On one stretch of Oxford Street that is 108 cafés within five minutes and 294 within ten.
  • The nearest station. Its name, its mode and the walk to it, on every one-address answer and in the Dossier.
  • Streets graded A+. Name a city and a trade, and the parades graded A+ for it come back as a list and on the map, spaced apart for a chain choosing its next sites.
  • Where the trade is still missing. Trestate counts the parades in a city where a trade is absent and the ground can carry it.
  • What traded here before. In the Paris region, the businesses on record at an address and around its parade, each with its dates, from the national business register. Three cafés are on record at 95 Rue de Charenton since 1 June 2004.
  • What businesses sold for. In the Paris region, the middle price and the middle 50% of business sales for the trade in each arrondissement or commune, from the BODACC register.
  • The Dossier. Printed pages for one address in the Paris region, with London's on the way: the grade, the station, the competition on the map and the address's record, ready for the bank, the accountant or the seller.
  • Boards to share. Save answers to a board and send it as a link. Someone without an invite sees a fixed snapshot of the board and an invitation to join.
Ask: “Where should I open a café in SW11?” The parades come back in grade order, the top one outlined on the map.
Ask in London: “Where should I open a café in UB2?” Ten parades in grade order, each named by its address.
The Dossier d’emplacement for 95 Rue de Charenton: the grade, the cafés within a five- and a ten-minute walk, and the sources and dates behind every figure.
03 / 06

Worth knowing before you negotiate

In the Paris region, where Trestate holds a decade of trading history, about a third of café and bar closures on record were not the end of the café. In four of five of those, a different company carried on the same trade at the same address.

Being told "the café closed" and being told "the café carried on under a different company" lead to very different offers. Every figure behind that line can be checked against the public business register.

From the Paris shop film on LinkedIn: the nearest stations and an A+ for cafés at the door. Scene simulated for illustration; the answer is Trestate’s own.
04 / 06

How it is built

Precompute the city, then answer in a read.

Each H3 res-9 hexagon carries a 112-feature Location Feature Vector, so a grade, a shortlist, a competition count and a Dossier all read the same row. A Python and DuckDB pipeline, run by Make, folds two countries of incompatible public data into one schema: Overture and OpenStreetMap for the streets and shops, ONS census on Ordnance Survey boundaries in London, INSEE census on IGN IRIS and the Sirene business register in the Paris region. features.yaml is the schema source of truth, code-generated into SQL, Go and TypeScript, and a trade is a YAML archetype, so a new one is a spec file, not code.

A Go API over PostgreSQL and PostGIS runs on Fly.io in London. The landing and the Ask app are React 19 on Cloudflare Pages behind a Worker, with Deck.gl and MapLibre for the maps. Every edge deploy runs the apps' test suites. While Ask is open, a real browser then opens the live app as a tester, and a blank map or a card that did not load closes Ask automatically until it is fixed.

The Trestate mark.
05 / 06

How I run it

Trestate runs as one owner and a company of role-scoped AI agent seats, from a CEO and product to research, sales and platform, with a Bar Raiser that reviews every API release before it ships. Each seat works from its own rulebook and register. I rule on what the seats escalate, and each ruling goes into a numbered decision register that outranks every seat.

The discipline is written down. Research pre-registers each hypothesis before it runs and records a null as a result. Every outward sentence comes from one copy source with its rules attached. When a correction lands, it is swept across every document that carried the wrong claim.

06 / 06

What's next

A domain and a finished brand identity replace trestate.pages.dev.

Invites move people from the waiting list into the private preview. London's graded area widens, Lyon and Manchester are extracted and waiting to load, and the takeover record built for cafés extends to other trades.

Decisions and why

  1. Precompute the city, then answer in a read.

    WhyThe grade, the shortlist, the competition count and the Dossier all hit the same vector. Normalising two countries of data happens once; every new question on top is nearly free.

  2. A live check closes the product, not just the build.

    WhyAfter an edge deploy, while Ask is open, a real browser opens the app as a tester. If a map renders blank or a card fails to load, Ask closes itself, so a tester never meets a broken answer.

  3. A street is a position, not a cell id.

    WhyBuyers ask about Oxford Street, not an H3 index. The product names the door and puts a pin on the parade.

  4. Archetypes as YAML.

    WhyA bakery and a household are specs. A new vertical starts as a spec file, not code.

Technical highlights

  • 01

    Go API over a 112-feature H3 res-9 vector: every question is a precomputed read

  • 02

    Ask: plain-language street and trade resolution in English and French, named positions, Gemini narration over served numbers

  • 03

    The Dossier: a PDF per address in the Paris region, with London's on the way, rendered from the API's page data by a FastAPI and Chrome service on Fly

  • 04

    Business sale prices from BODACC and nearest-station rows served beside the grade

  • 05

    features.yaml as the schema source of truth, codegen into SQL, Go, and TypeScript

  • 06

    Python + DuckDB pipeline (Make targets) across UK and French national datasets, including a decade of Sirene trading history

  • 07

    Cloudflare Pages + Worker edge and a Fly.io API in London, with deploy gates that open the live app as a tester

  • 08

    YAML archetypes: a new trade is a spec file, not code

Outcomes

By the numbers.

London
+ the Paris region
5 & 10 min
walks of competition, mapped
Visit project

See it in the wild.

This case study covers how Trestate was built — these links open what ships today: the live product, repo, demos, and supporting material.

Recent activity

The changelog.

October 2026
↗ Next case study

ScrollCollector

ScrollCollector collects the posts in your social feeds that match your own prompts, as you scroll, and compiles them into a diary you can filter and export. By default it decides on your device: a small sentence-embedding model bundled with the extension (System 1) scores every post, the browser's built-in model (Gemini Nano in Chrome, Phi-4-mini in Edge) is asked only about the posts System 1 is unsure of, and a keyword engine fills any gap. Cloud providers are off unless you turn one on. Widen the side panel and the stream reflows into columns.