The problem, then the work.

The problem. Their cluster had accumulated mixed GPU architectures across generations and vendors. Multi-node workloads assume uniform devices, so jobs that ran fine on a homogeneous partition failed or silently degraded once scheduled across the full estate.

The problem. The lab needed to study plasticity loss and temporal-difference instability in reinforcement learning agents. Neither effect isolates cleanly, because existing environments offer no controlled way to introduce noise or measure degradation consistently across runs.

The problem. Internal policy documents have to be checked against statute. Probabilistic retrieval returns false negatives, and a missed conflict between company policy and law is legal exposure rather than a defect to fix later. Separately, handbook content needed to reach employees as video, generated from documents already on the platform, without spending generation cost on prompts that were never going to produce anything usable.