Choudhary Usman Alam

Summary

What I did

• Fine-tuned 4+ Large Language Models (LLMs) on a therapy-specific dataset using PyTorch and Hugging Face
• Developed a minimalistic internal web interface using React and TypeScript to facilitate expert testing and chat logging
• Implemented an automated evaluation pipeline using DeepEval to benchmark inference latency and accuracy
• Managed iterative fine-tuning cycles with continuous hyperparameter optimization and automated evaluation feedback loops.

Results

• Reduced hallucination rate by 35% through targeted fine-tuning on therapy-specific datasets.
• Optimized model verbosity and response length to align with expert recommendations, significantly reducing irrelevant content generation.
• Improved response reliability by implementing advanced context-retention strategies for nuanced conversational flows.
• Established benchmarking for ROUGE scores and hallucination frequency using DeepEval.
• Collaborated with 4 domain experts to utilize curated longitudinal therapy datasets, ensuring high-fidelity context retention.
• Iteratively optimized hyperparameters through rigorous testing cycles, using DeepEval metrics to drive performance gains.
• Reduced hallucination rate by approximately 35%
• Improved response reliability through targeted fine-tuning