EN/ES
~/projects/sherpa

S.H.E.R.P.A

A casualty evacuation command screen for the front line, where drones have taken away the golden hour. It runs offline on a single laptop.

sheet
role
Team of 4 (TuringDB challenge): built Replay, the map's disruption view and the demo video, and pitched it
period
2026.10.02 → 2026.10.04
context
Built at European Defense Tech Hackathon. Situational Hub for Evacuation, Resupply and Patient Allocation
status
archived
code
public
stack
Python, YOLO, NVIDIA Parakeet, Sentence embeddings, TuringDB (Cypher), OpenStreetMap routing, React, TypeScript, deck.gl, Leaflet
results
Fuses drone detections and soldier triage reports into one graph and recommends a TCCC triage category; the medic decides in one click
Re-plans every affected evacuation route in the same minute a drone closes a road, fully offline on one laptop
Built at the European Defense Tech Hackathon in London; I built the Replay (a 251-minute run in 0.8 s) and pitched it
story

The problem

Getting wounded soldiers to surgery within an hour nearly halved combat deaths in Afghanistan. In Ukraine, cheap drones make every trip to the front dangerous, and medics can’t reach every casualty. The bottleneck is no longer speed alone: it’s deciding who goes first, by which vehicle, along which road, while the situation keeps changing.

What we built

Two data sources, one graph. A drone spots who is lying still, and a soldier beside the casualty answers simple triage questions in an offline phone app. Both reports are fused in a TuringDB graph, each with a position and a time.

Recommendation, human decision. The system recommends a triage category from TCCC doctrine, with its reasons and a confidence score that rises when a second source confirms the injury. The medical coordinator accepts it or changes it with one click.

Automatic logistics. Ambulances and ground robots (UGVs) are dispatched over real OpenStreetMap roads near Orikhiv, Ukraine.

Adapts as the world changes. When an enemy drone closes a road, every route that used it is re-planned in the same minute. When a robot’s battery dies with a casualty on board, another one is sent.

Voice questions, in English or Ukrainian, answered from the graph in about 100 ms. Fully offline.

Replay. Every simulated minute is a database commit, so the whole run can be stepped through to audit what the coordinator knew, and when.