AI Research Engineer | Generative AI | Computer Vision | Robot Learning
I build research-driven machine learning systems, from model design and large-scale training to efficient production deployment.
I build research-driven machine learning systems that move from model design and distributed training to optimized production inference.
I am an AI Research Engineer with 4+ years of experience across generative AI, computer vision, multimodal learning, and efficient inference. At Fynd, I develop image and video systems involving Flow Matching, diffusion models, super-resolution, model distillation, distributed PyTorch, and TensorRT.
Previously, at Wobot.ai, I built large-scale video analytics systems for detection, multi-object tracking, and cross-camera association. My current research connects generative vision with embodied intelligence: learning representations, trajectories, and policies that are accurate, efficient, and deployable.
| Flow Matching Super-Resolution Distilled into a single-step generator |
Lower Inference Latency TensorRT-optimized deployment |
Daily Active Users Production image and video services |
| Lower Video Latency Up to 52% lower compute cost |
Camera Deployments Across more than 250 locations |
StackCube-v1 Success Flow Matching Transformer + ViT policy |
High-resolution inference: 2500x2500 inputs processed in approximately 4-5 seconds on one NVIDIA L4 GPU.
Accepted at the ICML 2026 Workshop on Structured Probabilistic Inference and Generative Modeling (SPIGM).
A geometry-aware Flow Matching formulation for straighter generative transport paths and more efficient generation.
Sep 2023 - Present Β· Flow Matching super-resolution, single-step distillation, SDXL and FLUX training, video segmentation and inpainting, controllable generation, distributed PyTorch, and TensorRT deployment.
Feb 2022 - Sep 2023 Β· Production detection and tracking, multi-camera analytics, CPU and GPU pipeline optimization, Docker, and NVIDIA Triton deployment.
π€ mini-pi0Multimodal robot-learning framework using ManiSkill, MuJoCo, wrist-camera observations, and visuomotor Flow Matching policies. 95.5% StackCube-v1 success Β· VLA Β· IL Β· RL View benchmarks β |
π Flow-Based ModelsConditional Flow Matching experiments for image and video generation with optimal-transport paths and ODE sampling. 13+ stars Β· PyTorch Β· Generative Modeling |
ποΈ VideoGen MeanFlowComplete video-generation training and inference pipeline using MeanFlow with a Diffusion Transformer. Video Generation Β· DiT Β· Flow Models |
π Zero-DCETensorFlow implementation of zero-reference low-light image enhancement. 49+ stars Β· 8+ forks |
More applied ML work: π©Ί PraNet Polyp Segmentation Β· ποΈ TinyML Audio Classification
I am building an SO-ARM101 from individual components rather than using a pre-assembled system. The goal is to connect simulation-based policy development with physical data collection, evaluation, and learned-policy deployment. Hardware integration and real-world policy testing are currently in progress.
Representation Learning Multimodal Learning Distributed Training DDP FSDP
Diffusion Models Flow Matching SDXL FLUX Super-Resolution Flow Distillation One-Step Generation Image Generation Video Generation IP-Adapters Textual Inversion LoRA PEFT
Object Detection Multi-Object Tracking Image Segmentation Video Segmentation Image Restoration Video Restoration Vision-Language Models Controllable Generation
Vision-Language-Action Models Robotic Manipulation Visuomotor Policies Imitation Learning Reinforcement Learning Flow Matching Policies ManiSkill Robosuite Sim2Real
NVIDIA Triton Inference Server MLflow GPU Optimization Model Serving Data Curation Experiment Tracking
Bachelor of Science in Computer Science, University of Mumbai, 2022
CGPA: 9.21/10
I am interested in research and engineering conversations around generative vision, efficient deep learning, multimodal intelligence, and learning-based robotics.
- Email: mail2tauhidkhan@gmail.com
- LinkedIn: tauhid-khan-24bb45177
- Google Scholar: Tauhid Khan
- OpenReview: Isokinetic Flow Matching
Research depth. Engineering rigor. Systems that run outside the notebook.






