Taha Azzim

• Founded dailypapers, a newsletter service that sends a daily selection of top research papers along with summaries. Built on top of Gemini Agents for paper summarization and evaluation, hosted on OVHcloud, using Node.js, TypeScript, PostgreSQL, and SendGrid for email delivery.

• Implemented and deployed an End-to-End real time Computer Vision Collision Avoidance System based on drivable area segmentation and 3D tracking of Vulnerable Road Users (VRUs). The system runs on trucks and buses, it provides audio alerts and visual pictograms via displays to alert drivers in safety-critical scenarios. The system also replaced a previous vendor solution, saving on licensing fees and is currently deployed on the latest generation of Ford Trucks. • Helped develop a computer vision based parking assistance system for trucks that tracks trailer-end in real-time and provides zoomed visualization on in-cabin displays and ground-projected guidance lines for accurate maneuvering. • Implemented model optimizations pipelines, including structured pruning, INT8 quantization (QAT and PTQ), and custom operators which reduced model size by 85 percent and delivered a sevenfold increase in inference speed on Ambarella. • Developed and integrated neural networks on resource-constrained ECU modules (Ambarella CVFlow), which met the strict real-time processing requirements while under memory constraints. • Performed safety-critical analysis of model failures for road safety using advanced evaluation metrics. • Led a team of ten annotators and built automated labeling pipelines in CVAT using foundation models for auto-labeling and fine-tuned them to scale production dataset. • Managed the recording of large-scale real-world driving data in various weather conditions. • Setup and deployment of internal compute infrastructure: installation and configuration of 5 GPU servers, network switches, CUDA drivers, Docker environments, and secure SSH access within the internal network. • Established data pipelines and backup infrastructure using Google Cloud Storage Buckets, including implementation of dataset and annotation synchronization to ensure data traceability. • Trained Large Scale Models using Vertex AI Custom Training: DETR and SAM. • Deployed Teacher models on Vertex AI Endpoints to automate annotation of collected data and to improve the speed of large-scale data annotation processes.

• Trained stereo depth and segmentation models for robotic obstacle detection. The work included RAFT, DPT, DETR, YOLO, and SAM, which were optimized and tuned and merged for deployment on NVIDIA Jetson. • Developed a Python SDK that links cameras (Intel RealSense, ZED, and Luxonis) to compiled models (TensorRT, ONNX, CVFlow and OpenVINO) inside the embedded systems (NVIDIA Jetson, Orin, Ambarella, and Luxonis chips) and sends the outputs depending on client’s system. • Owned decisions for customer deployments. Each stack was customized to meet the costs, latency and hardware constraints. • Contributed to the development of a multi-camera depth estimation model for Parrot drones using fisheye lenses. Focused on optimizing the model and adapting training from synthetic data to real-world datasets. • Led code reviews for 9 developers to ensure company’s packages compatibility. • Developed Q&A metrics that translates complex technical data points into direct answers.