Ahamed H

Fine-tuned YOLOv8n on a 2,801-image PPE dataset containing 10 safety equipment classes; achieved mAP50: 0.702 and precision 0.877.
Developed a real-time safety monitoring pipeline for PPE compliance detection, automated violation identification, and severity classification.
Built and deployed a Dockerized Streamlit application on Google Cloud with CPU-only inference optimization, achieving approximately 45 ms/image inference latency.

Developed a hybrid retrieval pipeline integrating FAISS, BM25, Reciprocal Rank Fusion (RRF), and cross-encoder reranking to improve document retrieval relevance.
Built an automated evaluation framework using RAGAS to measure faithfulness, answer relevancy, context precision, and context recall for RAG system performance assessment.
Applied query reformulation and response validation techniques to enhance retrieval quality and reduce hallucinations in generated responses.

Built an AI-powered podcast question-answering system that retrieves YouTube transcripts, chunks content using overlapping sliding windows, and enables natural language interaction with podcast episodes.
Implemented lightweight keyword-based retrieval with term-frequency scoring and synonym expansion to identify relevant transcript segments without using embeddings or vector databases.
Developed grounded answer generation with Gemini 2.5 Flash, providing exact timestamps, direct transcript quotes, YouTube deep links, and graceful refusal for out-of-scope queries.