Zaid Ghazal

Blend Optimizer

BeyondAI · Petrochemical ML Platform

Built a platform for petrochemical companies (BP, Saudi Aramco) that trains ML models on historical lab data to predict product characteristics from ingredient blends, removing the need for costly physical trials. Delivered data cleaning, model training, and evaluation services end-to-end, no code required from users.

Impact: Replaced expensive, slow physical blend experiments with a self-serve ML platform covering the full clean → train → evaluate → infer workflow.

Computer Vision Fraud Detection

BeyondAI · Security Gate CV System

Designed a camera-only fraud detection system to replace expensive infrared sensors at security gates, using a centroid-based tracking algorithm to detect piggybacking, tailgating, crawling, and wrong-direction movement from live camera feeds.

Impact: Eliminated the need for additional hardware sensors by reusing existing CCTV infrastructure for real-time fraud detection.

GenAI Auto-Labeling Pipeline

BeyondAI · Vision-Language Models for Dataset Automation

Designed and deployed a generative AI auto-labeling service using vision-language models (Qwen), served with vLLM and Ollama, to automatically generate class labels for image data, accelerating dataset preparation for in-house YOLO models.

Impact: Significantly reduced manual annotation effort, letting the team iterate on detection models faster.

GenAI RAG Knowledge Base Chatbot

BeyondAI · Azure

Led the retrieval and re-ranking components of a retrieval-augmented generation (RAG) pipeline for a client's internal support knowledge base, built on Azure Data Factory and Azure AI Search, replacing slow manual searches with a generative AI chatbot interface.

Impact: Measurably improved retrieval quality (MRR, Hit Rate@k) and end-to-end response accuracy over the legacy manual search process.

Physical AI and Intelligent Systems Researcher

University of Michigan

• Built a PID controller autotuning framework and an LLM-in-the-loop framework that generated PID parameters from system feedback, validated in both simulation and on physical robots.
• Contributed to Maison, a University of Michigan Innovation Partnerships initiative researching secure LLM systems for safe generative AI adoption.
• Researched Vision-Language-Action (VLA) models and World Models for robot environment reasoning and decision-making.

Impact: Published the PID autotuning research in a peer-reviewed conference on control systems and robotics.

Rapid Proof-of-Concept Delivery

BeyondAI · Computer Vision, GenAI & VLA Applications

Delivered fast proofs-of-concept across computer vision, generative AI, and vision-language-action (VLA) applications to validate feasibility and demonstrate value to clients, including real-time people tracking, foot-traffic counting, and behavior detection in school environments, moving from research to working demo in short cycles.

Impact: Consistently turned research ideas into client-facing demos fast, a core strength in a startup environment.

Research Mentor

RISE Research

Mentor high school students 1-on-1 through original robotics research, guiding them from an initial idea through methodology, analysis, and writing, toward submission to peer-reviewed journals or conferences.

Impact: Paying forward the guidance I received through FLL and WRO to the next generation of student researchers.