I train diffusion transformers for faces and video, and make models fast enough for real-time production.
- Diffusion Transformers: a 303M-parameter flow-matching DiT that swaps facial identity in video while keeping pose, expression and gaze, trained on the MareNostrum 5 supercomputer.
- Real-time video: a GAN-based, landmark-free face anonymization model distilled from a slower teacher, and the video pipeline that runs it in production.
- Inference optimization: PyTorch FP32 → TensorRT FP16, latency 49 → 14 ms (3.5x), throughput 20 → 70 FPS.
- Video diffusion research: probing LTX-2.5 (22B) and temporal video DiTs with ablations, attention-head analysis and CKA.
My work code is private; these are my public projects:
| Project | What it shows |
|---|---|
| camera-motion-compensated-tracking | OC-SORT multi-object tracker written from the paper, with global (camera) motion compensation. Checked against the official code on MOT17: +2.42 HOTA on moving-camera sequences. Tests and CI |
| language-model-pretraining | Small GPT pre-trained from scratch on TinyStories: my own byte-level BPE tokenizer, RoPE, RMSNorm, SwiGLU and a KV cache (3.5x faster generation). 0.570 bits per byte after 58 minutes on an M2 laptop. Tests and CI |
| fast-fourier-transform-quality | FFT metrics for lost detail, flicker and keyframe pulsing in images and video, shown on real photos and x264 encodes. Tests and CI |
| brain-tumor-detection-mri | MSc project: YOLOv5 detection at 97.2% mAP (binary) and 94.6% mAP (4-class), CNN classifier, Grad-CAM |
| transformer-from-scratch | Decoder-style Transformer from first principles in PyTorch, with attention visualization |
| dcgan-from-scratch | DCGAN on CIFAR-10 with resumable training and FID / IS evaluation (FID 43.96) |
| Image-Understanding-Photogrammetry | MSc course work on camera geometry: keypoints, matching, epipolar and multi-view geometry |
| Camera-Trackpad | Real-time hand-gesture mouse with MediaPipe and OpenCV |
PyTorch · Diffusion / Flow Matching · Transformers · LLM pre-training · GANs · TensorRT · ONNX · CUDA · PyTorch DDP · SLURM · OpenCV · YOLO · Multi-object tracking · MediaPipe · Python · C++ · pytest · GitHub Actions