Image

Salman Naqvi

fastai Forums Twitter Artstation Email

AI researcher bridging theory with practice in vision-language models, embodied AI systems, and diffusion models. Strong background in distributed training, model optimization, and leveraging AI tools to enhance development workflows. Currently pursuing B.Sc. in AI & EdTech at EdUHK as a full scholarship recipient.

GPU Algorithm Optimization

  • Optimized Meanshift clustering algorithm achieving 6x speedup (1.5s → 0.25s) through batch processing and other GPU acceleration techniques
  • Read more

DiffEdit Implementation

  • Implemented DiffEdit paper from scratch, building a complete text-based image editing pipeline with Stable Diffusion
  • Successfully generated segmentation masks and identified VAE compression artifacts in latent space
  • Read more

PyTorch Reimplementation

  • Reimplemented PyTorch core neural network components from scratch (layers, backpropagation, autograd, optimizers, dataloaders)
  • Progressively rebuilt with PyTorch to understand modular architecture
  • Read more

Stable Diffusion Implementation

  • Built custom Stable Diffusion pipeline from components (CLIP, U-Net, VAE)
  • Implemented classifier-free guidance and negative prompts
  • Read more

Backpropagation Tutorial

  • Authored technical blog post deriving backpropagation from first principles using plain language
  • Read more

VLM Researcher (Project Lead) | Jun. 2025 – Aug. 2025

Focus Tech Lab (2025 HK Techathon Gold Medalist)

  • Diagnosed core bottleneck in robotic arm pipeline: VLM’s inability to interpret 3D spatial relationships
  • Integrated RoboBrain 2.0, saving weeks of development time
  • Built attention heatmap visualization tool to diagnose VLM failures
  • Developed attention-steering framework to manipulate VLM focus at inference time

Embodied AI Researcher | Oct. 2024 – Jun. 2025

Multimodal Intelligence Laboratory, The Education University of Hong Kong

  • Fine-tuned Qwen2.5VL-3B on RGB-D images, improving spatial reasoning accuracy by ~6%
  • Identified fundamental limitations in VLMs for spatial intelligence tasks
  • Findings directly informed subsequent spatial intelligence work at Focus Tech Lab
  • Analyzed state-of-the-art methods in robot manipulation (under Dr. Fu Hong)

AI Engineer (Project Lead) | Jul. 2024 – Aug. 2024

Keweya Education Technology Limited – HKSTP Incubatee

  • Built a career visualization app using PhotoMaker V2 and Stable Diffusion XL, allowing students to see themselves in future careers
  • Integrated face embedding for personalized image generation with LoRA modules
  • Developed a publicly accessible web-hosted full-stack Gradio application with story generation mode

LLM Researcher (Project Co-Lead) | Jun. 2024 – Jul. 2024

READily Limited – HKSTP Incubatee

  • Fine-tuned 70B+ parameter LLMs to generate tone-constrained Cantonese lyrics (novel task)
  • Implemented distributed training on 8-GPU cluster using FSDP and QDoRA
  • Reverse-engineered BPE tokenizer and developed custom layer hooks for tonal control

Programming Languages: Python

ML/AI Frameworks: PyTorch, Hugging Face (transformers, diffusers, datasets, tokenizers, peft, accelerate), fastai

Tools & Libraries: Pandas, NumPy, Scikit-Learn, Gradio, Jupyter, nbdev, FastHTML Quarto

Other: 3D Graphics (Blender), Technical Writing, Teaching

Languages:

  • English: Native
  • Mandarin: Conversational proficiency, can read technical documentation (approx. new HSK 5)
  • Urdu: Fluent

Image

Feb. 21st 2025 – Meet Salman Naqvi: the bright spark in AI and EdTech | See what the university’s written about me!

Dec. 27, 2024 – Featured among “The EdUHK Rising Stars” - EdUHK 30th Anniversary | See the university’s mini documentary on me!

Jun. 12, 2024 – Hong Kong SAR Government Scholarship Fund Reaching Out Award

Dec. 7, 2023 — Invited Speaker at the Advanced Study Institute (ASI) Online Symposium | The Education University of Hong Kong and Hong Kong Baptist University

Apr. 2023 – Nov. 2023 — Co-lead fast.ai Part 2 Online Study Group | From Deep Learning Foundations to Stable Diffusion | Discord

Mar. 17, 2023 — KUAS-E Scholarship Offer | Kyoto University of Advanced Sciences, Japan

Feb. 24, 2023 — EdUHK Full Entrance Scholarship | The Education University of Hong Kong

Nov. 2022 – Jun. 2023 – Built AI community “The Dataspace” - 1100+ members | Discord

The Education University of Hong Kong

B.Sc. (Honors) in Artificial Intelligence and Educational Technology

Expected Graduation: May 2027

  • EdUHK Full Entrance Scholarship (Feb 2023)

IELTS – 8.5/9.0





Back to top