Mohammad Yakub

At Humanoid Lab, IIIT Dharwad, I contributed to AI-based speech processing research by designing, implementing, and evaluating machine learning pipelines. I developed an automated audio data pipeline using yt-dlp and pydub, then trained a Retrieval-based Voice Conversion model that achieved an F0 RMSE of 39.54 Hz.

I developed a CI/CD-style LLM evaluation pipeline that tests prompt changes against a 72-case golden dataset, detects regressions and model drift, and can block merges when critical regressions exceed configured thresholds. I also implemented baseline comparison, LLM-as-a-Judge evaluation, HTML reporting, and Slack alerts.

I built a RAG pipeline combining dense vector search with BM25 retrieval and Reciprocal Rank Fusion, with cross-encoder reranking and grounded responses. For my automatic braking system project, I designed and evaluated Random Forest and neural network models that achieved 75% accuracy in collision prediction.