autodebug
Agentic debugger that reproduces, bisects, root-causes, and fixes real bugs from a repository and a bug report.
AgentsLLMsPython
AI systems,
built around your product.
I design, build, and optimize production-ready AI systems — from agentic applications and custom LLMs to scalable machine learning infrastructure.
About
From research labs to production AI.
I'm Shayan — an AI engineer focused on turning modern machine learning research into products people actually use. I grew up in a village in Iran and followed a path from electrical engineering into deep learning, ML research, and production AI.
Building software isn't only my job — I genuinely enjoy it. Outside of work I keep exploring emerging AI technologies so my toolkit stays current with where the field is heading.
Research-backed0→1 builderAgentic systemsEnd-to-end delivery
Education & Experience
Resume.
A concise view of my academic and professional journey.
- EducationB.Sc., University of TehranElectrical & Computer Engineering; early work on autonomous vehicles and intelligent transportation systems.M.Sc., University of TehranAdmitted as an exceptional-talent candidate; shifted into deep learning, with a thesis on pedestrian detection for autonomous vehicles.PhD, University of WaterlooMachine learning research, publications at top-tier ML venues, a one-year ML engineering internship, and industry collaborations.
- Professional ExperienceAI Engineer, Huawei (2025 – present)Building agentic systems for compiler applications such as code optimization, and leading projects in collaboration with academic partners.
Core Expertise
Agentic Systems
LLM Development
Fine-Tuning
Retrieval-Augmented Generation
AI Performance Optimization
ML Engineering & Deployment
Services
How I can help you ship AI.
Focused engagements across the modern AI stack — from prototype to production system.
AI & LLM Development
Design and build intelligent applications powered by large language models, including AI assistants, agentic systems, and domain-specific AI products.
LLM Fine-Tuning & Customization
Adapt open-source language models to custom datasets and use cases with techniques such as LoRA, PEFT, and distributed training.
AI Performance Optimization
Improve inference speed, reduce latency, and lower infrastructure costs using GPU optimization, Triton kernels, quantization, and efficient model serving.
Retrieval-Augmented Generation
Build AI systems that use organizational knowledge bases through vector databases, retrieval pipelines, and large language models.
ML Engineering & Deployment
Develop production-ready machine learning pipelines, APIs, deployment systems, and scalable AI infrastructure.
AI Consulting
Provide guidance on AI strategy, model selection, system architecture, technical feasibility, and production best practices.
Selected projects
A sample of what I build.
Open-source and research projects from my GitHub. More on the way.
GCRL — Generative Causal Representation Learning
Reference implementation of a causal representation learning method based on variational autoencoders, tied to the ICML 2023 paper on out-of-distribution motion forecasting.
Causal MLVAEPyTorch
DeepFusion
Deep-learning pipeline for pedestrian detection in autonomous vehicles, from the pedestrian-detection line of work in the master's thesis.
Computer VisionDetectionPython
Cuda_CNN
MNIST classification implemented directly in CUDA — a low-level look at how convolutional networks map onto GPU kernels.
CUDAGPUCNN
Research & publications
Peer-reviewed work behind the practice.
Selected publications on causal representation learning, LLMs for code, computer vision, and intelligent transportation.
- Findings of the Association for Computational Linguistics (ACL)·2026
PerfCoder: Large Language Models for Interpretable Code Performance Optimization
J. Yang, S. Lu, H. Liu, S. Shirahmad Gale Bagi, Z. Fazel, T. Czajkowski, D. NiuUses LLMs to suggest and explain performance optimizations in source code, so the recommendations are auditable rather than black-box.
- arXiv:2602.06142·2026
Protean Compiler: An Agile Framework to Drive Fine-grain Phase Ordering
A. H. Ashouri, S. Shirahmad Gale Bagi, K. Satheeskumar, T. Srikanth, J. Zhao, I. Saidoun, et al.A framework for fine-grain compiler phase ordering that improves the quality of generated code across workloads.
- arXiv:2402.11124·2024
Implicit Causal Representation Learning via Switchable Mechanisms
S. Shirahmad Gale Bagi, Z. Gharaee, O. Schulte, M. CrowleyA follow-up on causal representation learning that models interventions as switchable mechanisms in the underlying data-generating process.
- International Conference on Machine Learning (ICML)·2023
Generative Causal Representation Learning for Out-of-Distribution Motion Forecasting
S. Shirahmad Gale Bagi, Z. Gharaee, O. Schulte, M. CrowleyA causal representation learning approach for motion forecasting that stays reliable when test data shifts away from the training distribution.
Contact
Have an AI project in mind?
Tell me what you are building, where the technical challenge is, and what stage the project is currently in.