---
title: "FlowTwin: we won the Google DeepMind track"
description: "Our team, Low Cortisol, took first place in the Google DeepMind track at RAISE Summit 2026. I helped shape the project, designed with agents and operated the demo."
date: 2026-10-19
lang: en
translationKey: "flowtwin"
translationStatus: machine
kind: project
name: "FlowTwin"
period: "RAISE Summit 2026"
role: "Low Cortisol team member; direction, design with agents and demonstration"
status: "Team hackathon prototype"
tags: ["AI", "hackathon", "teamwork"]
related: []
sourceUrls: []
canonical: "https://shipulin.md/en/projects/flowtwin/"
---

At RAISE Summit 2026, our team Low Cortisol won the Google DeepMind track with FlowTwin, a demonstration model of patient flow through a hospital. We had started as a team of five with an idea for a furniture builder. After talking to the mentors, we dropped it. By the end of the day, we were discussing something else entirely: making patient flow and delays inside a hospital visible.

Slava had invited me. Gabriel proposed the new direction after our furniture idea failed to excite the mentors. Letting go of my idea stung. But “AI for a hospital” still sounded like a collection of big words. We needed an angle: what could we actually show at a hackathon?

## Find something real

We started looking for a hospital publishing information in real time. Hong Kong had open, aggregated data: waiting times at 18 hospitals, updated every 15 minutes.

The project now had a connection to reality. We could build the demonstration around actual waiting times, rather than inventing both the hospital and its problems.

My agents spent all night designing the interface. In the morning, we discovered they had forgotten to connect the data. Slava hooked it up on the fly. By morning, we had the interface. The data connection had some catching up to do.

## A queue has a history

We built FlowTwin as a demonstration of an operational digital twin. It shows the model hospital's state changing over time: who is waiting, where a delay has emerged, and how it affects the next steps. In the demonstration, you can follow one patient's scenario, introduce a delay and see the model change.

During the first presentation, we were asked rather pointedly why this needed AI at all. I found the question useful. We had to explain more precisely that the agent needed a history of events, rather than just the latest snapshot of a queue. Where had the person already been? What were they waiting for now? Which constraint had appeared earlier?

After the first round, the team was deflated. It was hot, the presentation had been interrupted, and the whole thing felt unpromising. But we made it through. Before the final, we shortened the story and rebuilt the demonstration. I operated the demo and answered a question.

Low Cortisol took first place in the Google DeepMind track at RAISE Summit 2026. The team was Slava, Gabriel, Laiba, Manel and me.

## Where the demo ends

The pressure on the hospitals shown in the demo comes from real, aggregated waiting times. But the individual patients, their names and their journeys are synthetic. A change in waiting time after an action in the interface is the behaviour of the demonstration model, not a reduction in waiting at a real hospital.

The repository separates those layers: real data, synthetic characters, numerical assumptions and model calls. The prototype deals with time, queues and handoffs. It does not diagnose people or choose treatments.

Before the final, we had to shorten the explanation to what we could actually demonstrate. “Why does this need AI?” had been a useful rehearsal.

## Sources

- [My account of the hackathon, pivot and presentations](https://t.me/humanagentinteraction/213)
- [The project and team](https://www.linkedin.com/posts/alexeishipulin_we-came-to-the-largest-ai-hackathon-in-ugcPost-7480105651952758785-lHRn/)
- [FlowTwin repository and data provenance](https://github.com/shipaleks/raise-hackathon-flowtwin)
- [FlowTwin demo](https://divine-roch-grounded-bf796e9a.koyeb.app/)