At a San Francisco hackathon on 26 September, 55 teams built working projects around Jev, a model from TypeSafe AI, using GMI Cloud’s inference service for text, image, video and voice generation. The useful idea was to give quick, typed decisions to one model and leave the actual writing, images and audio to others.
GMI Cloud Watch analysis
What happened
GMI Cloud says more than 500 people RSVP’d and teams had three hours to build. Jev can return a choice, score or yes/no answer in a fixed format, with a probability attached; it does not generate prose. The pattern demonstrated at the event was to use Jev for branching decisions, then send generative tasks to models on GMI Cloud’s MaaS platform through an OpenAI-compatible endpoint.
The winning project, Jevolution, used Jev to decide the next moves of simulated rabbits and wolves, while a larger model set up the experiment. Other teams used the split for tasks including stroke-scale assessment and coding-agent command checks. These were hackathon projects, not evidence of clinical or production deployment.
Why it matters
Many AI agents ask a generative model both what to do and how to say it. Separating those jobs could make some workflows quicker and reduce unnecessary generation calls, while letting developers choose a different model for each step. At the event, teams also swapped models by changing a string, according to GMI Cloud.
That is a useful design pattern, not a measured performance result. The account offers no comparative timings, cost figures or independent evaluation of the projects. A brisk demo is a promising start; it is not the same thing as a dependable system outside the room.
Our read
The interesting contribution is the architecture: use a constrained decision call where software needs a branch, and reserve generative models for work that produces something people will read, hear or see. That is more specific than the usual promise to put an agent in charge of everything, which is generally how the paperwork starts breeding.
Developers can take the pattern away and test it on a small, measurable task. The event’s results are encouraging examples, but teams should compare accuracy, latency and cost against their existing approach before giving the split a permanent desk.
What to watch
- Whether TypeSafe publishes technical details and independent evaluations of Jev’s decisions.
- Whether developers can reproduce the reported speed and reduced generation-call pattern.
- How the approach handles uncertain decisions in workflows with real-world consequences.
Discussion spark: For an AI agent, should routine branching decisions go to a fast, constrained model while generative models handle the output, or does splitting the work create more failure points than it removes?
Sources and evidence
- Hacking on Jev: notes from JEVATHON (1 October 2026, 15:00 UTC)
not affiliated with or endorsed by GMI Cloud