~/omorros
ProjectsWorkJournal
Back

Hackathon · {Tech: Europe} × VEED Hackathon · London 2026

Lantern

Know the building before you go through the door.

A dispatch tool that briefs firefighters on a burning building before they arrive, from its old property listing.

View code on GitHub
Event
{Tech: Europe} × VEED Hackathon, London
Result
Overall Winner and Best Use of fal
Stack
Python, FastAPI, Next.js, Playwright, OpenAI, fal
Team
Mykyta Yakivets and Bill
The three Lantern teammates with their arms around each other under the VEED neon sign

Why it matters

Firefighters often go into a burning building blind: no floor plan, no idea where the fire started or which room the person they are looking for is in. Yet the outside of almost every UK building is on Street View, and the inside of many homes was photographed for an old property listing. Lantern joins the two during the 999 call. Size-up is the fire service word for the quick read of a building on arrival; we moved it to before the truck leaves.

How it works

Speech to text

The caller, as it happens

OpenAI · gpt-4o-mini-transcribe

"The kitchen's on fire."

GLiNER2

Finds the address, fire and casualty

Pioneer · every fragment

Casualty: back bedroom

Outside

Reads the street and the plot

Google Maps · gpt-5

Door on the left, no rear access

Listing agent

Finds the house in old sold listings

Holo 3.1 · Playwright

  • listing found
  • floor plan pulled

12 photos and a floor plan

Room graph

Every room, its doors, its photos

gpt-5 points · code draws

8 rooms, 3 ways in

Route

From the kerb to the casualty

along the room graph

Front door, 5 waypoints

Crew brief

Each line marked by its source

VEED and Kling on fal

30s brief ready

The callThe buildingThe briefSolid boxes are what we built. Dashed ones are outside services.

Decisions we made

  1. The model points, code draws the rooms

    gpt-5 reads the floor plan and points at each room label on a grid. The rooms themselves are worked out without it: the walls are thickened until the door gaps close, then each room is filled outwards from its label.

    Trade-off. More code than asking the model for boxes, and the wall thickness is one fixed number. When the fill fails, we fall back to the box the model drew.

  2. Every line says where it came from

    Each line of the crew brief is marked call, street, listing or plan: what the caller said this minute, what Street View shows, what an old listing shows, or our own inference from those. Listing and plan lines are shown fainter.

    Trade-off. A listing can be years out of date, so the layout, one of the most useful parts of the brief, always arrives marked as the least certain.

  3. A saved result behind every stage

    We picked five sold houses, ran each one through the whole live pipeline on the day, and saved every stage. If a service fails or a key is missing during the demo, that stage replays its saved result and the brief still arrives.

    Trade-off. Those five houses are all the agent was tested on, so we have no success rate to quote.

Photos

Presenting Lantern beside a screen showing photos of a building's roomsThe team building Lantern around a table of laptops and drinksWith Abai, a friend from H Company who was at the hackathon, both giving a thumbs upWorking on Lantern on a laptop during the hackathon

The full engineering is on GitHub.