Project case study
Factory N+1
shippedA factory that redesigns itself. Nemotron agents propose each floor layout; a frozen simulator decides what ships. Won NVIDIA Best Use of Nemotron.
Factory N+1 treats every change to a factory floor as a revision. Nemotron agents propose a layout and predict its score. A frozen simulator runs 100 packages through it. The revision ships only if it beats the current floor by at least 1% on 9 validation days and no goal check gets worse.
Results
- Rev 11 cut average simulator penalty from 18,061 to 3,283, 82% lower.
- 100/100 shipped, 0 damaged, 0 mis-sorts on all 15 development days.
- 54 model calls and $0.042 for the winning run. $0.355 of inference across every run.
- A LoRA-tuned Nemotron Nano 8B optimizer cut mean penalty 12.64% on 20 unseen factories, against 0.14% for the base model and 4.27% for random legal moves.
System design
- Nemotron 3.5 Super VL: designs layouts through typed tool calls and predicts the score.
- Nemotron 3.5 Lightning: research on You.com and isolated draft tests.
- Crusoe Managed Inference: serves every call and caches the shared-memory prefix.
- Frozen simulator: hash-pinned, the only grader. Agents never score themselves.
- Per-edit scoring: each change is scored alone, so lessons cite measured effects.
- Three.js replay: the original and improved floors on one clock.
Current boundary
The floor is synthetic and the results are simulated. The 82% headline comes from validation days; the locked test split stayed closed. The fine-tuned optimizer is a narrow pilot: 21 of 60 proposals were invalid and no factory reached the full goal. Nothing here is a real-world throughput claim.