Public procurement compliance LLM
Reads a public bid notice and decides, for each of 24 legal-violation items, whether the notice breaks the rule, quoting the exact sentence that proves it.
Fixed model (gemma-4-26B-A4B-it), no fine-tuning, no internet, and two hours of inference for the whole test set.
- vLLM pipeline on one L40S that groups related items into a few prompts per notice and checks that every quote is a verbatim substring of the notice.
- Per-item decisions that combine the model's judgment with rules from the statute package. On dev, removing a false positive was worth about 4.5 times more F1 than catching a miss, so each item's gate is tuned for precision.
- Self-labeled a sample of the 20,000 unlabeled notices with frontier LLMs.
- Paired bootstrap tests on dev to separate real gains from run-to-run noise. The result advanced the team to the final evaluation round.
Python, vLLM, gemma-4-26B-A4B-it