Habeeb Olawale Hammed

○ Led the migration of a production human action recognition system from tree-based models to Transformer architectures, enabling more accurate temporal modeling of 3D pose sequences for real-time embodied learning. ○ Increased per-class precision by 118% and recall by 22% across 17 action categories, significantly improving the reliability of movement detection used in live classroom gameplay. ○ Built scalable, GPU-optimized data pipelines for large-scale video and pose datasets, cutting preprocessing and training time by 35% and accelerating model iteration cycles. ○ Designed and deployed an LLM-powered AI assistant that converts gameplay logs into real-time student and class performance summaries, feedback generation, reducing teacher analytics workload by over 60% . ○ Developed a teacher-facing dashboard (React, FastAPI) with visualizations and alerts, reducing manual log analysis time by >50% . ○ Integrated ML models into production systems in collaboration with product and engineering teams, ensuring low-latency inference, monitoring, and smooth rollout to end users.

○ Built an LLM-powered clinical intelligence pipeline to extract structured medical data and generate treatment recommendations from unstructured cancer reports, boosting extraction accuracy by 40% . ○ Designed, implemented, and evaluated a Retrieval-Augmented Generation (RAG) system using FAISS and dense embeddings, achieving 85% precision and 90% recall on real clinical queries. ○ Improved document layout and section understanding using Mixtral 8x7B, increasing information retrieval accuracy by 30% on complex multi-page medical records. ○ Developed evaluation and benchmarking pipelines to validate model performance, robustness, and hallucination risk before expe rimental deployment in research workflows.

○ Built an automated deep learning pipeline to classify over 100,000 satellite images, supporting large -scale Earth observation and environmental monitoring. ○ Improved classification accuracy by 25% using transfer learning with pre-trained CNNs, enabling more reliable detection of land and atmospheric patterns. ○ Optimized data loading and inference pipelines, cutting end-to-end processing time by 35% and reducing compute and memory overhead for large imagery batches.