At Turing, I compared AI model responses for accuracy, relevance, completeness, and instruction following, maintaining an approximately 94% approval rate on submitted work. I also designed tests for edge conditions and invalid inputs.
In a later Turing contract, I developed and evaluated system prompts exceeding 10,000 words and tested agent function-calling across multi-turn conversations. I documented model failure scenarios, including missed constraints and incorrect tool selection.
I built a secure digital evidence storage prototype using AES-GCM and PBKDF2, and an edge-to-cloud license plate recognition pipeline using ESP32-CAM, YOLO, and OCR. I also trained wine quality prediction models, achieving a test MSE of approximately 0.538.
