Case Studies

#Eliminating Hallucinations in Legal Document Review

Enterprise Legal Firm // 99.9% Citation Accuracy | 40x Faster Review

Architected an evaluator-first RAG pipeline that enforces strict provenance on all generated claims. By moving from a pure generation model to a generation-evaluation loop, we achieved deterministic reliability.

The Problem Space

The client was relying on standard zero-shot LLM prompting to extract clauses from 500-page legal contracts. While extraction speed was high, the model suffered a 4% hallucination rate—unacceptable in compliance environments where a single fabricated clause can invalidate a case.

Architectural Shift: The Evaluator Loop

Instead of relying on prompt engineering to suppress hallucinations, I decoupled the architecture into two distinct agents: a Generator and a programmatic Judge. The Generator creates the initial extraction. The Judge then runs a deterministic substring-matching algorithm (and a secondary constrained LLM check) to verify that every single claim generated exists verbatim in the source context.

Outcome

By enforcing this structural provenance, we eliminated hallucinated claims entirely. The system now processes 10,000+ pages a day with 99.9% verifiable citation accuracy, fundamentally shifting the firm's document review economics.

#Real-Time Object Detection for Remote Sensing

AgriTech Startup // Sub-50ms Inference | 92% mAP

Designed and deployed a custom computer vision architecture optimized for edge devices, enabling real-time crop analysis from drone feeds without relying on cloud inference.

The Problem Space

Drones capturing remote agricultural data were bottlenecked by cloud dependency. Pushing gigabytes of raw video feed over intermittent 4G networks caused massive latency, preventing real-time decision making for automated crop dusting.

Edge-First Vision Architecture

I spearheaded the transition to edge inference. I architected a custom YOLO-based detection pipeline, quantized the models using TensorRT, and optimized the inference graph for deployment directly on Jetson Nano edge modules attached to the drones. I also implemented a robust annotation strategy to fine-tune the model specifically for top-down canopy topologies.

Outcome

The drones now run inference locally in sub-50ms per frame. The system achieves 92% mAP without a continuous internet connection, allowing the client to scale their fleet globally into zero-connectivity rural zones.