Muhammad Ahmed

AI Engineer

Pioneer Corporation, Tokyo, Japan

Sep 2023 - Present • 2 yrs 8 mos

Engineered a bilingual (English/Japanese) real-time AI driving assistant on Android: Featuring an end-to-end on-device ASR and TTS pipeline; implemented a neural network router to dynamically classify and switch user queries between a super-lightweight edge model and a large cloud LLM for the execution of 70+ driving commands, achieving 70%+ overall accuracy. ● Lightweight GAN-Inspired Intent–Slot Model for Edge Android Control: Designed a compact, GAN-inspired audio language model for on-device intent–slot parsing—under 50 MB and sub-200 ms end-to-end latency—while matching heavyweight audio-LLM accuracy for real-time Android automation. ● Audio LLM-Based Android Agent & Automation: Built a voice-driven Android agent using MERaLiON for ASR intent decoding and ADB-shell hooks for hands-free control, with context retention, retry logic, and robust error handling—achieving over 90 % command success across 50+ real-world scenarios. ● Multimodal Vision-Language & Video AI Systems: Fine-tuned, integrated, and quantized six VLMs (Chat-UniVi, LLaVA 1.5, CogVLM, MobileVLM 2, VideoLlama, Florence-2) with advanced prompt engineering to halve real-time inference latency. Developed a unified video–GPS sync pipeline with grid-based UI, automated PDF/slide reporting, and online/offline RAG (ensemble + miniGPT v2) to boost long-form query accuracy on diverse driving datasets. ● Traffic-Rule Classification & Accident QA: Architected a hierarchical VLM classifier encoding 400+ Japanese traffic rules into modular pipelines, generating automated demo videos for compliance checks. Extended with a multimodal car-to-car accident QA system that answers natural-language queries on real incident footage. ● Advanced Detection, Tracking & OCR Frameworks: Replaced Faster-RCNN with zero- and one-shot deepAnything→YOLOX pipelines—cutting latency from ~20 s to 1–2 s—and added DeepSORT for automatic license-plate blurring. Built lane-tracking and Japanese traffic-sign OCR (OpenCV + VLM) and fused

AI Research Engineer - Computer Vision and Machine Learning

Retrocausal, Redmond, WA, USA

Jun 2022 - Sep 2023 • 1 yr 3 mos

● Conducted computer vision research and solved problems involving classification, object detection, and segmentation under Dr. Quoc-Huy Tran's supervision. ● Contributed to research papers on self-supervised learning for video activity recognition, achieving a 2% performance improvement and outperforming state-of-the-art methods in 2 publications. ● Explored video data generation in projects using latent diffusion models and conducted research on action recognition utilizing 2D human pose estimation, surpassing state-of-the-art methods by a 10% margin. ● Implemented a zero-shot unsupervised video segmentation algorithm in a demo product to be presented to Apple and Siemens. ● Developed a Video based Large Language Model (LLM).

Data Scientist

Pikky, India

Nov 2021 - May 2022 • 6 mos

● Made a patented food recommendation engine from scratch and deployed it to real-time users. ● Developed a novel way of recommendation for food utilizing machine learning, natural language processing and computer vision techniques with accuracy of above 90%. ● Worked with cross-functional teams: product managers, developers, and designers, to efficiently deliver and integrate data-driven solutions, adapting to shifting requirements and priorities in an Agile framework.