File 04 · Hacktoberfest Rabat · UM6P · Best use of open models track · 24 Oct 2026
FleetSense
“Salam chef, lmoteur skhoun bzaf.” Four agents read a driver’s Darija message with the fault codes, and return one decision with its cost.
- ~30 sper breakdown, running liveSimulated
- 4 + 12agents and fleet tools on open weightsMeasured
- 8,000MAD penalty avoided in the main scenarioSimulated
A truck on the A7 is overheating with 10 pallets of medicine on board. Someone has 90 minutes to get another truck there, or the company pays an 8,000 MAD penalty.
The agents work through the fleet’s own tools. Which part failed? Water pump, with the evidence listed. Repair or replace? Replace: the part is at nearly twice its rated life. Which truck takes over? No spare can make it in time, but T-03 is already heading to Marrakech with 14 free pallets.
The fleet manager gets one card and one button. Every tool call, result and approval goes on an audit trail.
What it refuses to doActing alone or doing the maths. The model reasons and chooses, code recomputes every cost, ETA and penalty, and nothing happens until the fleet manager approves.
How it runs
- ReadDriver message in Darija plus the vehicle’s fault codes.
- DiagnoseWhich part failed, with evidence, on the open-weights DeepSeek model.
- PlanRepair or replace, which supplier, which depot, which truck takes the load.
- RecomputeCode recalculates every cost, ETA and penalty the model proposed.
- ApproveOne card, one button, and a full audit trail.
Honest caveatThe fleet, suppliers and prices are simulated, and the team wrote the scenarios. A working prototype, not a deployed product.
Next fileWolf Station →