Cairo · EG
Antony Saleeb
Full-Stack & Applied-AI
Available for roles & collaborations
Fig. 01 — Flagship
DeepFract
Encode path
01 / 05
Input
Attention
Residual
Quad-tree
Encoded
Source frame
Classical fractal compression is slow because the encoder searches an enormous space of block self-similarities. DeepFract splits that job across specialised networks that each handle one part of the decision, then uses quad-tree partitioning to spend detail only where the image actually needs it.
Hard parts
- 01Coordinating multiple specialized models so they improve ratio and quality without blowing up encode time.
- 02Designing attention that surfaces structural self-similarities useful for fractal block matching.
- 03Bridging a PyTorch inference backend with a responsive Flutter mobile client.
Stack
Partition
One cell per job — the same cut the encoder makes.
- 03
PyTorch
Models
- 01
Flutter
Client
- 02
FastAPI
Inference
- 04
Python
Runtime
- 05
Computer Vision
Perception
Methodology & results
Evaluated on rate–distortion — compression ratio against PSNR — versus classical fractal and transform-coding baselines. The benchmark set is being re-verified against a fixed test corpus before the headline numbers go up here.
Next.jsReactTypeScriptTailwind CSSPythonFastAPINode.jsREST APIsFlutterDartFirebasePyTorchComputer VisionAttention modelsGitGitHubSupabaseNext.jsReactTypeScriptTailwind CSSPythonFastAPINode.jsREST APIsFlutterDartFirebasePyTorchComputer VisionAttention modelsGitGitHubSupabase