Hasnain Syed
Vibe coding since 2020.
Projects
Social media scheduling API that lets AI agents publish videos, images, and carousels across multiple platforms.
AI-powered study platform where students upload notes, slides, and textbooks to get source-cited flashcards, quizzes, summaries, and answers.
AI interior design tool that transforms room photos in seconds with virtual staging, sketch-to-render flows, recoloring, and multiple style presets.
Prompt-to-UI product for generating polished mobile app screens fast, helping teams iterate from idea to build-ready designs.
Lightroom-inspired photo editor that runs entirely in the browser, built with Rust compiled to WebAssembly for near-native image processing and React for the UI.
Optimizations
FLUX.2-klein-4B Inference
RTX 5090
1.90s → 0.48s
- Reduced text-to-image latency from 1.90s to ~0.48s on an RTX 5090 using FP8 quantization (torchao) and torch.compile — a ~75% speedup at 1024x1024, 4 steps.
Ideogram-4 Inference
H100
14.80s → 3.99s
- Cut Ideogram-4 generation from 14.80s to 3.99s on an H100 by adding flash attention, dequantizing nf4 to bf16, and compiling with torch.compile — a 73% speedup at 1024x1024, 12 steps.
Experience
Research Assistant
Anglia Ruskin University
Jan 2025 - Apr 2025
- Built a novel multi-platform misinformation dataset by extracting data from TikTok, X.com, and Odysse.
- Engineered data pipelines to structure and prepare unstructured text for NLP model training.
- Created interactive dashboards to visualize insights and communicate findings to stakeholders.
Publications
Semantic Segmentation For Landslide Detection Using Segformer
SGAI-2024 · Lecture Notes in Computer Science
This research aims to achieve state-of-the-art results for landslide segmentation using aerial images. Experimental results show the advantage of the Segformer model in segmenting landslide areas, with the largest Segformer variant achieving an Intersection over Union (IoU) score of 87.795% on the Unmanned aerial vehicle (UAV) dataset, surpassing the previous state-of-the-art model, Multiscale Feature Fusion and Enhancement Network (MFFENet), by 3.4%. On the Satellite (SAT) dataset, Segformer attained an IoU score of 79.300%, outperforming the previous best model, DeepLabv3+, by 11.163%. For the combined UAV&SAT dataset, Segformer achieved an IoU score of 85.157%, surpassing DeepLabv3+, the best previous model by 5.032%.