Co-Founder & Systems/AI Architect @ Hoosha AI π§
Computer Engineering @ University of Tehran & Sharif University of Technology
I am an AI Researcher & High-Performance Systems Engineer based at the University of Tehran (UT) and cross-institutional Teaching Assistant at Sharif University of Technology (SUT).
My work bridges the gap between Deep Generative Modeling (Flow Matching, Score-Based Diffusion, VAEs), LLM Reasoning & Alignment (GRPO, SFT, Task Arithmetic), and Low-Level Systems Infrastructure (Kaleido Distributed GPU Engine, C/xv6 Operating System Kernels, Go Servers).
- π Location: Tehran, Iran (
UTC +3:30) - π University: Computer Engineering @ University of Tehran (UT)
- π« TA Affiliation: Compiler Construction @ Sharif University of Technology (SUT)
- πΌ Leadership: Co-Founder & Systems/AI Architect @ Hoosha AI π§
- π§ Primary Email:
tahamajlesi@ut.ac.ir| Secondary Email:Tahamajlesice@gmail.com - π Personal Website: https://tahamajs.github.io/
| Domain | Technologies & Frameworks |
|---|---|
| AI & Deep Learning | PyTorch CUDA Transformers Flow Matching GRPO Diffusers Scikit-Learn |
| Low-Level Systems | C C++ Go xv6 Unix Kernel ARM Assembly Compilers Make |
| Backend & Web Systems | Django Django REST Framework PostgreSQL Docker Celery Redis Git |
| Data & Research | NumPy Pandas SciPy Matplotlib Seaborn Jupyter LaTeX |
Continuous probability paths, Flow Matching, Score-Based Diffusion (DDPM/SDE), RealNVP Normalizing Flows, and Vision-Language Models (VLMs like PaliGemma).
Group Relative Policy Optimization (GRPO) + SFT for 4B LLMs on GSM8K mathematical reasoning, Task Arithmetic shortcut patchers, and ReAct agent workflows.
Selective Amnesia algorithms for targeted class-forgetting and unlearning in Conditional Variational Autoencoders (VAEs).
- π Unlearning from VAE
Kaleido: First-principles distributed LLM training engine jointly optimizing 4 training dimensions across GPU compute nodes.
- π Kaleido Engine
Operating system kernel extensions, custom system calls, and process scheduling in the xv6 Unix OS (C language), plus hybrid Java compiler construction.
π University of Tehran & Sharif Coursework Directory (18 Repositories)
| Course Name | Institution | Primary Stack | Repository Link |
|---|---|---|---|
| Deep Generative Models | π University of Tehran | PyTorch, CUDA | Deep_Generative_models_course |
| Natural Language Processing | π Sharif University | PyTorch, Transformers | Natural-Language-Processing-Projects |
| Machine Learning (M.Sc.) | π University of Tehran | Python, Scikit-Learn | machine_learning_course |
| Operating Systems Lab (xv6) | π University of Tehran | C, Assembly | xv6-Modified_OS_Lab_Projects |
| Computer Networks | π University of Tehran | Python, Sockets | Computer_Network_Projects |
| Deep Reinforcement Learning | π Sharif University | PyTorch, Gym | Deep_Reinforcement_Learning_SUT |
| Trustworthy AI | π University of Tehran | PyTorch, Safety | Trustworthy_AI_Course_Assignments |
| LLMs Final Project | π University of Tehran | PyTorch, Qwen | LLM_UT_FINAL_PROJECT |
| Data Science | π University of Tehran | Pandas, Seaborn | Data_Science_Cource_Assignments |
| Software Engineering II | π University of Tehran | Python, Django | SE2_UT |
| AI Ethics | π University of Tehran | Ethics, Bias | Ethics_In_AI_Course |
| Advanced Programming (Turtix) | π University of Tehran | C++, SFML | Turtix_Game |
| Advanced Programming | π University of Tehran | C++ | Advanced_programming_Projects |
| Artificial Intelligence | π University of Tehran | Python, Search | AI-CA1_-Agents_And_Search_Algorithms |
| Computer Architecture Lab | π University of Tehran | Verilog, ARM | Computer_Architecture_LAB |
- π Personal Website: tahamajs.github.io
- π§ Hoosha AI Substack: hooshaai.substack.com
- π€ HuggingFace Profile: huggingface.co/tahamajs
- π¬ Telegram: @tahamajlesii
Designed & Maintained by Mohammad Taha Majlesi β’ University of Tehran & Sharif University of Technology





