At UNC PRATON, I own the technical scope of a production RAG assistant for a 1C-Connect support line. I built its evaluation harness across seven LLM-as-a-judge metrics, improving correctness from 2.92 to 4.16 and context relevance from 3.05 to 4.38.
I designed retrieval root-cause diagnostics that reduced failure cases from 97 to 4, and built a generator for 924 validated evaluation questions. I also shipped hybrid retrieval on Qdrant for a 116K-chunk corpus.
On a freelance receipt-digit OCR project, I built a 0.3M-parameter CNN with CTC loss that achieved 98.05% accuracy on 564 hand-labelled images. I shipped it with a self-check against golden reference images.
At FastDeploy, I reworked a production OCR pipeline for documents of arbitrary length and multi-column layouts, and built real-time sports video analytics using Python, OpenCV and YOLOv8. At Daily Challenge, I led three developers and took two projects to corporate rollout, including an LLM post-analysis system that cut manual review from about three hours to 15 minutes a day.
