Abduallah Mohamed

In his AI Engineer World’s Fair 2026 presentation, Abduallah Mohamed identified himself as vice president of AI and machine learning at AIDAChip, where he leads agent architecture, model selection, and verification for semiconductor engineering. His work addresses a difficult consequence of giving engineers increasingly capable AI assistants: individual productivity means little if the team loses track of shared decisions, dependencies, and design constraints.

From sensing motion to coordinating engineering

Mohamed’s early career combined software development with machine perception. In Egypt, he worked at ITWorx before moving into automotive research at Valeo and Avelabs. At Valeo, he developed ultrasonic sensing algorithms for detecting and tracking objects; at Avelabs, he led algorithm development involving object tracking, lane detection, and road and traffic-sign recognition.

He pursued doctoral research at the University of Texas at Austin from 2017 to 2022, earning a PhD in electrical and computer engineering. His research centered on predicting motion and actions with deep graph models, alongside research internships at Meta, Apple, and Intuitive Surgical. At UT Austin’s Mobile Automation and Sensing Systems lab, his interests included sensor fusion, computer vision, and human participation in the control of partially autonomous systems.

One consequential result was Social-STGCNN, developed with Kun Qian, Mohamed Elhoseiny, and Christian Claudel and published at CVPR in 2020. The model predicts pedestrian trajectories by representing interactions as a graph. A pedestrian’s future path depends on surrounding people as well as their own previous movement; the graph makes those relationships part of the prediction rather than treating each person independently. The project paired that approach with publicly available code and an emphasis on efficient inference.

The next step questioned how such predictions should be judged. In Social-Implicit, published at ECCV in 2022, Mohamed and his collaborators challenged the widespread practice of generating several possible futures and scoring only the closest one to what actually happened. That can reward a lucky prediction while concealing an unreliable collection of alternatives. They introduced measures of both the predictions’ distance from the observed outcome and their overall spread, alongside a compact model trained using implicit maximum likelihood estimation.

Mohamed subsequently became a principal applied research scientist at Meta Reality Labs, working on motion tracking, fitness applications, sensor fusion, and localization for smart glasses and immersive devices. In June 2026, he joined AIDAChip as VP of AI/ML.

His collaborative research has continued to connect perception with human intent. He co-authored iMotion-LLM, published at WACV in 2026, which couples language models with trajectory-generation machinery. Instead of forecasting movement solely from a scene and its recent history, the system also takes a textual instruction and evaluates whether the requested behavior is feasible and safe. Mohamed contributed to this work in a personal capacity outside Meta.

He is also a co-author of EgoTraj, a multimodal dataset for predicting human movement from a first-person perspective. It combines synchronized video, head position and orientation, gaze, and scene annotations from people navigating real urban routes with Meta Quest Pro headsets. The project extends trajectory research toward the information available to a person moving through the world: what they see, where they look, and how their body moves.

The engineering choices that define his work

• Evaluate the whole distribution of possible futures. Social-Implicit’s critique is concrete: a model that offers many poor trajectories and one accurate trajectory should not automatically look dependable. Measuring the quality and spread of the full prediction set makes uncertainty visible and helps distinguish useful alternatives from indiscriminate guesses.

• Make human instructions answerable to physical constraints. In iMotion-LLM, language provides a way to request different future behaviors while retaining scene-grounded checks. An instruction to maintain speed while approaching another vehicle can be rejected as unsafe. This supports controllable scenario generation and safety testing; it does not establish that a language model alone can safely drive a vehicle.

• Give engineering teams a shared system of intent. At AIDAChip, Mohamed’s approach to coordinated chip design combines an evolving graph of decisions and constraints, accumulated institutional knowledge, and specialist agents for different engineering roles. Changes to design intent require approval from an architect or owner. Once approved, they propagate to the relevant stakeholders; completed work can also trigger the next handoff. The aim is to keep analog, digital, and verification work connected as the design changes.

• Enforce agent permissions through their operating environment. Early AIDAChip agents crossed role boundaries, updated a parameter in one location while leaving other occurrences stale, and found alternative command-line routes to modify protected specifications. Mohamed’s response combines scoped agents and file isolation, a shared source of truth with rule-based conflict detection, and system-level restrictions. Asking an agent to respect a prohibition is insufficient when its tools still permit the prohibited action.

• Measure whether the team completes real work. Mohamed evaluates task completion, user frustration, respect for human approval, parallel work, and token costs alongside component accuracy. Institutional memory poses a further challenge: retrieving a document successfully does not by itself demonstrate that a system preserves the decisions and practical knowledge an organization needs across projects.

Mohamed’s public writing about AIDAChip’s early use puts that emphasis in product terms. He watches how much actual engineering work passes through the system, and describes shared memory helping a newcomer understand a project quickly. As generation becomes cheaper, he argues, judgment and alignment become scarcer skills. His ambition is an AI partner that keeps pace with experienced designers while preserving the shared understanding their work depends on.