30 internal teams want to fine-tune open-weight models on their own data — some with LoRA adapters that finish in under an hour, some with full fine-tunes that run for two days. Today everyone emails the ML infra team and waits.
Build a self-serve platform: a team submits a job, gets GPU time from a shared 40-GPU cluster, and ends up with a versioned, evaluated model artifact they can promote to serving without infra involvement.
Teams' training data must never be visible to other teams. A single team's jobs should never be able to starve everyone else's queue, and a training run that dies six hours in should not lose six hours of work.
Build the architecture on a canvas: place the components, configure them, connect them into a data flow, and write a short reason for each one. The AI reviewer grades your design against a rubric written specifically for this problem.
A team's job is preempted 30 hours into a 48-hour full fine-tune to make room for another team's job. Trace exactly what state is saved, and what the job does when it resumes.
What does your automated eval gate actually check before an adapter is promotable, and what stops a team from gaming it?
How do you prevent one team from monopolising the cluster by submitting fifty jobs at once, without a human ever intervening?
Minimum 5 components · needs a wide desktop screen