Legal systems produce large volumes of high-stakes decisions in unstructured natural language, making large-scale empirical analysis costly, difficult to reproduce, and unevenly accessible. This bottleneck is especially acute for legal analytics and policy evaluation in low-resource languages such as Portuguese. To address it, we present a resource-efficient pipeline for information extraction from Brazilian criminal case law that reuses a legacy dataset to fine-tune open-weight LLMs with Q-LoRA. Operating in a small-data setting and using schema-constrained JSON generation, the pipeline extracts 47 legal variables spanning charges, evidence, and sentencing outcome. In held-out evaluation, a fine-tuned Phi-4 (14B) model achieves 92.8% accuracy and 0.826 macro-F1, approaching proprietary baselines while retaining the cost and privacy benefits of local deployment. We then use the extracted data in a case study of the short-term effects of a recent Brazilian Supreme Court ruling on drug decriminalization, finding no statistically significant change in trafficking-conviction rates (p≥0.05), a pattern consistent with short-run institutional inertia. More broadly, the paper contributes a reproducible framework for legal NLP and shows how legacy empirical datasets can support scalable legal analytics under severe resource constraints.