Tammam Alhadwah

AI-based monitoring system for early wildfire and agricultural threat detection. Built a tiny detection transformer with a MobileNet backbone and a hybrid custom transformer encoder, benchmarked against a fine-tuned YOLOv8-Nano on a 21K-image dataset of fire and smoke instances.

Tuned aggressively for recall, applied Exponentially Moving Average smoothing to stabilise per-frame predictions, and deployed with TensorRT on NVIDIA GPUs for real-time inference.

End-to-end automated turret that tracks and follows moving targets for laser-tag gameplay. YOLO11 detection runs against a phone-camera video feed with real-time multi-object tracking, exponential smoothing for movement, and a manual-override keyboard layer.

Motion handled by an Arduino motor-control system driving relay actuators, with TensorRT-optimised inference for responsive targeting.

Therapeutic conversational agent for multi-turn dialogue in Arabic. Rather than fine-tuning an existing model, the dialogue architecture was built from scratch — trained with PyTorch Lightning and tuned for empathetic, contextually aware responses in a therapeutic setting.

Classifier for LiDAR-like 3D point-cloud representations of furniture. Implemented the PointNet architecture in Keras to handle irregular point-cloud data directly, without voxelisation.

Cleaned a noisy and mislabeled dataset of ~1,690 social-media comments and messages, then fine-tuned DistilBERT from Hugging Face Transformers to classify cyberbullying and harassment.