Open to senior ML / CV roles · France & Luxembourg

Computer Vision Engineer
with a PhD in computer science & AI

I build and deploy computer vision and vision-language systems that work in production. My background spans six years of applied research at INRIA and the University of Luxembourg, covering video understanding, pose estimation, multimodal learning, and efficient model design with PEFT adapters. I recently shipped a full MLOps pipeline serving Qwen2-VL on serverless GPU infrastructure with automated CI/CD, experiment tracking, and a public demo. PhD expertise in action detection, pose estimation, anomaly detection, and vision-language modeling. Marie Curie Fellow. Now bringing research depth to production AI engineering.

View my work → ↓ Download CV Google Scholar GitHub
Thionville, France (near Luxembourg) +33 751 890 606 aali.abidkhan@gmail.com
Abid Ali
Postdoctoral Researcher · CVI² Lab
SnT, University of Luxembourg
Google Scholar LinkedIn GitHub University profile Download CV
6+
peer-reviewed papers
incl. WACV, ICCV, AVSS
7+
years building deep
learning systems
3×
keypoint improvement on
pose estimation (TALON)
EU
Marie Curie Fellow
+ 2× hackathon winner
Full Stack
MLOps pipeline: serving, CI/CD, monitoring, serverless GPU deployment
Tech stack
Deep Learning & CV
PyTorch TensorFlow Vision Transformers CNNs / 3D CNNs PEFT / Adapters OpenCV SAM DINO LLMs / VLMs Qwen · LLaVA-Next Keras
Video & Perception
Action Detection Behavior Analysis ByteTrack DeepSORT Optical Flow Temporal Modelling 6-DoF Pose Est. Anomaly Detection Contrastive Learning Self-supervised (MAE)
Engineering & Infra
Python C / C++ MATLAB Linux Git SLURM AWS Multi-GPU (DDP) Docker Modal FastAPI GitHub Actions vLLM Weights & Biases CI/CD pipelines REST API design Serverless GPU NumPy · Pandas Scikit-learn Dataset Curation
Featured projects
TALON — Spacecraft Pose Estimation
Preparing · BMVC 2026
🛰️

Token-aligned lightweight adapters injected into vision foundation models for 6-DoF spacecraft pose estimation. Achieves 3.3× improvement in keypoint localization using less than 5% additional parameters. Delivered 31% reduction in pose error on SPADES and 2.5× lower error on SPARK.

PyTorch ViT PEFT 6-DoF PnP
Preprint coming soon
VLM Action Captioning Pipeline
Live Demo
🎬

End-to-end production MLOps pipeline deploying Qwen2-VL-2B-Instruct for video action understanding. Built the full stack from model serving to CI/CD automation: FastAPI inference server, Docker containerization, serverless GPU deployment on Modal, automated GitHub Actions pipeline, Weights and Biases experiment tracking, and public Gradio demo on Hugging Face Spaces.

PyTorch FastAPI Docker Modal GitHub Actions Weights & Biases Gradio vLLM
Live Demo API Docs GitHub
AdaTAD++ — Action Detection at Scale
ICCV 2025 · IEEE/CVF
🎬

Scaling temporal action detection with transformer-enhanced spatial-temporal adaptation. Co-first-author work introducing a new adaptation paradigm for long untrimmed video understanding on large-scale benchmarks.

Video Understanding Transformers Temporal Adapters
ICCV 2025
Autism Behavior Analysis in Video
WACV 2025 · IEEE/CVF
🧠

End-to-end system for recognizing social interactions and severity estimation in real-world therapy sessions. Combines weakly-supervised learning with large-scale video pipelines using YOLO, DeepSORT, and SAM to handle occlusions and noisy clinical environments.

Weak Supervision YOLO DeepSORT SAM
WACV 2025 · CBMI 2024
3D Gaze Estimation from Face Images
MS Thesis · Sejong University
👁️

Dual-channel CNN for unconstrained 3D gaze estimation from natural face images with real-time webcam tracking. Designed for low-cost accessible setups — no specialized eye-tracking hardware required.

CNN Gaze Estimation Object Detection Real-time
Sejong University 2020
Experience
Aug 2025 – Present
Luxembourg
Postdoctoral Researcher
CVI² Lab, SnT — University of Luxembourg
  • Designing spatiotemporal adapters for vision foundation models achieving 3.3× keypoint improvement with <5% added parameters.
  • Built real-time structured perception pipelines integrating foundation models with PnP solvers; bridged research into production-ready codebases.
  • 31% pose score reduction on SPADES · 2.5× lower error on SPARK benchmark via zero-shot domain adaptation.
  • Collaborated with aerospace experts at Zero-G lab; applied CV models to real aerospace environments.
  • Partnered with INRIA to integrate LLMs (Qwen, LLaVA-Next) for enhanced video analysis and contextual understanding in ASD research.
  • Investigated multimodality foundation models for identity-based deepfake detection with POST Luxembourg.
  • Designed multimodality models combining 3D CNN and Transformer networks across Skeleton, RGB, and text modalities.
  • Designed and deployed a serverless VLM inference pipeline serving Qwen2-VL-2B-Instruct on Modal GPU infrastructure with FastAPI, Docker, and automated CI/CD via GitHub Actions.
  • Built end-to-end MLOps pipeline covering model serving, containerization, experiment tracking with Weights and Biases, and continuous deployment to Hugging Face Spaces.
Nov 2020 – Jul 2025
Sophia Antipolis, France
Computer Vision Researcher (PhD)
STARS Team — INRIA France · Marie Curie Fellow
  • PEFT modeling: developed AM Flow adapters achieving 90% reduction in training time and 10× faster convergence with SOTA performance on SSv2 and Kinetics-400.
  • Scaled input resolution 120× with AdaTAD++ PEFT architecture, setting a new benchmark in long-duration video understanding.
  • Contrastive learning: investigated SigLIP, CLIP, and X-CLIP for video learning to identify unusual autistic behaviors in children.
  • Built end-to-end action detection systems for long, untrimmed real-world video streams.
  • Engineered large-scale preprocessing pipelines: YOLO, Faster R-CNN, DeepSORT, ByteTrack, SAM.
  • Built a 300-hour multimodal dataset with clinicians, advancing research on autistic children's behaviors.
  • Worked on vision-language models to enhance recognition and anticipation of multimodal actions.
🏆 Marie Curie Co-Funded Fellowship · 2020–2024
2023 – 2024
Martigny, Switzerland
Research Visit
Idiap Research Institute
  • Improved autism severity estimation accuracy by 30% through a weakly supervised video learning framework, reducing dependence on dense temporal annotations.
  • Self-supervised pre-training with SimCLR, MAE, and VideoMAE to reduce annotation needs and improve generalization in long medical behavior videos.
Feb 2018 – May 2020
Seoul, South Korea
Research Assistant
HCI Lab — Sejong University
  • Investigated 3DCNNs for object detection and tracking in videos.
  • Developed multi-stream CNN for 3D gaze estimation from unconstrained face images.
  • Improved user accessibility by 15% by creating a low-cost real-time eye-tracking system using object detection.
  • Designed and implemented human detection and tracking algorithms for a smart city surveillance project in Korea.
  • Investigated and optimized Faster R-CNN fire detection on real-world videos, improving detection reliability in safety systems.
🎓 Professor Scholarship · 2018–2020
Selected publications Full list on Scholar →
In prep
TALON: Token-Aligned Lightweight Adapters for 6-DoF Spacecraft Pose Estimation
Ali, A., Rathinam A., Aouada D.
Preparing to submit · BMVC 2026
Co-1st
Scaling Action Detection: AdaTAD++ with Transformer-Enhanced Temporal-Spatial Adaptation
Agrawal T., Ali A. (equal contribution), et al.
ICCV 2025 · IEEE/CVF International Conference on Computer Vision
Co-1st
Are Attention Maps Richer than we Imagined for Action Recognition?
Agrawal T., Ali A. (equal contribution), et al.
AVSS 2025 · IEEE Advanced Video and Signal-based Surveillance
WACV
Loose Social-Interaction Recognition in Real-World Therapy Scenarios
Ali, A., et al.
WACV 2025 · IEEE/CVF Winter Conference on Applications of Computer Vision
WACV
P-Age: Pexels Dataset for Robust Spatio-Temporal Apparent Age Classification
Ali, A., et al.
WACV 2025 · IEEE/CVF Winter Conference on Applications of Computer Vision
Education
PhD · Computer Science & AI
Université Côte d'Azur · INRIA
2020–2024 · Valbonne, France
"Video Analysis using Deep Neural Networks: An Application for Autism" — Advisor: Dr. François Brémond
MS · Computer Engineering
Sejong University
2020 · Seoul, South Korea
"3D Gaze Estimation from Natural Face Images using a Dual Channel CNN Network" — Advisor: Prof. Yong-Guk Kim
BS · Electrical Engineering
COMSATS University Islamabad
2012–2016 · Pakistan
"Real-time orange crop disease detection using computer vision and embedded systems"
Awards & certifications
🥇
1st place — 3IA Hackathon (Caranx Medical)
2023
🥉
3rd place — Varuna AI Hackathon
2022
🇪🇺
Marie Curie Co-Funded Fellowship (PhD)
2020–2024 · EU Horizon
📜
Computer Vision Nanodegree · Udacity & Deep Learning Specialization · Coursera
Industry certifications

Let's build something together.

I am targeting applied ML engineering and senior computer vision roles where I can develop reliable, high-performance vision systems from research through production deployment. Open to remote, hybrid, and relocation within France and Luxembourg.

email aali.abidkhan@gmail.com
linkedin linkedin.com/in/abid-ali-9404
github github.com/abalikhan
google scholar Full publication list
resume Download CV (PDF)