Ahsan Umar_

Ahsan Umar

AI/ML Engineer & Researcher

$ whoami --contact
email
ahsanumar.ai@gmail.com
github
https://github.com/codewithdark-git
linkedin
https://linkedin.com/in/codewithdark-in
huggingface
https://huggingface.co/codewithdark
orcid
https://orcid.org/my-orcid?orcid=0009-0005-8823-2686
location
Peshawar, Pakistan
  • 150+GitHub stars
  • 4research papers
  • 50+open-source projects
  • 4certifications

Experienced in Generative AI, Large Language Models (LLMs), agentic AI systems, NLP, and computer vision, with a background spanning production-grade AI pipelines, transformer fine-tuning (QLoRA/PEFT), and open-source tooling that has drawn 150+ GitHub stars. This work spans full-time and contract engineering, academic research coordination, technical mentoring, multimodal systems research, and supervision of an AI research lab.

Education. BS in Artificial Intelligence, Islamia College University Peshawar (Oct 2023 – present). Coursework includes Machine Learning, Deep Learning, NLP, Computer Vision, Data Structures & Algorithms, Linear Algebra, Probability & Statistics, and Software Engineering. Affiliated researcher at the Smart Technology Lab, with ongoing work on transformer models and medical-imaging publications.

Certifications. Deep Learning Specialization (DeepLearning.AI, 2024) · Generative AI with Large Language Models (AWS, 2024) · TensorFlow Developer Certificate (Google, 2024) · NLP with Transformers (Hugging Face, 2024).

Experience

AI Engineer — Contract

AlphaTechLogics

Mar 2026 – Present

  • Advises and drives advanced Generative AI R&D for production-grade LLM deployment across client verticals.
  • Implemented specialized AI consulting services in multimodal system design, RAG optimization, and agentic workflow automation, reducing AI feature deployment time by 30%.
  • Works cross-functionally with product and engineering teams to define AI roadmaps and evaluate emerging model technologies.

AI Engineer — Full-Time

AlphaTechLogics

Feb 2025 – Feb 2026

  • Developed FaseehGPT, a small transformer model for Arabic text generation, improving morphological understanding and inference speed by 40% in low-resource settings.
  • Created independent agentic AI systems with LangGraph and CrewAI to automate client SaaS workflows, reducing manual effort by 50%.
  • Led end-to-end LLM fine-tuning projects (QLoRA/PEFT) for enterprise clients and integrated domain-specific models into FastAPI architectures.
  • Designed scalable backend services for AI agents, optimized inference latency, and integrated solutions with client software ecosystems.

Research Lab Coordinator — Smart Technology Lab

Islamia College University Peshawar

Jan 2026 – Present

  • Coordinates the Smart Technology Lab research team and aligns the project portfolio with institutional and industrial research goals.
  • Sets best practices for experiment tracking, model versioning, and reproducibility with W&B, DVC, and Hugging Face Hub.
  • Facilitates collaboration among faculty, student researchers, and industry partners to convert research into products and publications.
  • Conducted Generative AI, LLM, and Computer Vision seminars, paper reading sessions, and hands-on sessions for 30+ participants.

AI Research Engineer

Islamia College University Peshawar

Jan 2025 – Dec 2025

  • Researches computer vision, NLP, LLMs, multimodal learning, and efficient transformer training paradigms for non-English languages.
  • Wrote two research papers under review on transformer optimization for low-resource languages and transformer models for medical imaging.
  • Supervised undergraduate AI research projects from conception to implementation in computer vision and NLP.
  • Co-developed algorithms for active research projects with faculty.

Blogs

12 entries · view all →

Long-form writing on models, tooling, and applied research.

  • › LeWorldModel: Stable End-to-End Joint Embedding Predictive Architecture from Pixels LeWM learns end-to-end world models from pixels using JEPA with a single SIGReg regularizer — no reconstruction loss, no pretrained features, no temporal smoothness loss. Achieves 96% success on Push-T (vs 78% PLDM, 81% DINO-WM) with ViT-S predictor. Key insight: isotropic Gaussian latent constraint (SIGReg) replaces multiple regularizers; temporal straightening emerges implicitly. Jul 19, 2026
  • › VL-JEPA: Joint Embedding Predictive Architecture for Vision-Language VL-JEPA learns a joint vision-language embedding space using only JEPA-style latent prediction + SIGReg regularization — no contrastive loss (CLIP), no generative loss (LLaVA), no reconstruction (MAE). Competitive zero-shot on ImageNet, Flickr30K/COCO retrieval, GQA/VQAv2 VQA. Jul 19, 2026
  • › Attention Is All You Need — Paper Breakdown The Transformer replaces RNNs/CNNs with pure attention. Encoder-decoder stacks of self-attention + FFN achieve SOTA translation (28.4 BLEU EN-DE, 41.8 EN-FR) at fraction of training cost. Introduces multi-head attention, scaled dot-product attention, sinusoidal positional encoding. Jul 15, 2026

Projects

3 entries · view all →

Open-source tools and applied systems.

  • › QuantLLM A lightweight toolkit for quantizing, fine-tuning, and deploying LLMs on consumer hardware. Mar 1, 2025
  • › DocsGPT A retrieval-augmented assistant for searching and exploring technical documentation. Feb 1, 2025
  • › DarkGPT A GPT chat explorer for real-time conversations across multiple LLM providers. Aug 1, 2024