JevBOT Rx delivers STAT medication through busy hospital corridors. Nurses task it by text. Code guarantees safety, and Jev supplies the split-second judgment an LLM is too slow to give.
60-second demo. The nurse ⇄ robot texting runs at 1.5×; the race is sped up 5×. Left: Jev. Right: DeepSeek-V4-Flash on GMI Cloud. Same floor, same people, same safety brake.
No kiosk, no app. Jev reads every message in one pass, with three typed questions: intent (delivery, status, cancel, confirm), urgency (routine, urgent, STAT) and destination. The robot replies with an ETA and texts again when it arrives.
The robot's brain is split in two. Code does everything that must be exact. Jev makes the one judgment call code can't: of these safe paths, which one, right now, given the people around me?
Jev returns a probability over the options it's offered in one pass. An LLM writes thinking tokens and then JSON. By the time its answer lands 2–5 s later, people have moved, so the robot waits or acts on a stale view.
Jev charges $0.042 per million input tokens, and output is free. LLMs pay for input, output and hidden reasoning tokens on every call.
Collision prediction, filtering, validation and the safety brake are deterministic and identical for both brains. The model only picks among paths already checked.
Planner: JevPilot's sampling-based local planner (similar to the Dynamic Window Approach) plus a shortest-path route over the corridor graph. Every question sent to Jev: jev-questions.md · Architecture PDF
The same 20 real robot decisions were sent to every model, with the same inputs and each LLM on its fastest setting. The full delivery race and stress test were run against DeepSeek-V4-Flash.
| Jev (TypeSafe) | DeepSeek-V4-Flash | GPT-6 luna | GPT-6 sol | GPT-6 astra | |
|---|---|---|---|---|---|
| Time per decision (p50) | 143 ms | 2,275 ms | 2,302 ms | 2,349 ms | 3,387 ms |
| Time per decision (p95) | 373 ms | 5,080 ms | 3,842 ms | 4,503 ms | 9,494 ms |
| Cost per decision | $0.000074 | $0.000197 | $0.000129 | $0.002262 | $0.013206 |
| Valid answers | 20/20 | 18/20 | 20/20 | 18/20 | 20/20 |
| STAT delivery, Pharmacy → ICU bed 4 | 157 s | 232 s | — | — | — |
| Stress-test collisions (brake off) | 0 of 2 runs | 2 of 2 runs | — | — | — |
Measured on seed 42 with 400 people and 30 carts. These are single-scenario runs, not a statistical study. In the stress test, both robots have the safety brake off and a patient steps into the lane about 2 s ahead. Stress-test video · Full-length race · Texting demo
You need a TypeSafe API key for Jev and a GMI Cloud key for the LLM brain.