Foundation models, transformers & AI systems

Mohsen Hosseini.

PhD, Computer Science, Machine Learning Scientist

I design and build transformer-based models, foundation models, and retrieval-augmented AI systems. My work focuses on representation learning, efficient model adaptation, and scalable generative AI, from structured scientific data to language.

5+
Years in ML research
6
Peer-reviewed publications
61
LLM book chapters
3
AI focus areas

Machine Learning Scientist with a PhD in Computer Science and over five years of research experience in deep learning, transformers, and large-scale representation learning. I specialize in designing and training neural architectures for complex data, from foundation models to retrieval-based AI systems.

My work spans generative AI, LLM systems, and scientific machine learning. I build complete ML systems, including model architectures, training pipelines, retrieval systems, and production tooling using PyTorch, JAX/Flax, Hugging Face, Docker, and HPC infrastructure.

Currently at York University, my research explores two directions: building world models for complex systems and developing efficient approaches for adapting foundation models using techniques such as LoRA fine-tuning.

Based inWaterloo, Ontario
Current roleAI Researcher, York University
PhDWestern University
MScUniversity of Tehran
GitHub github.com/mohsenh17
LinkedIn LinkedIn
Scholar Google Scholar

Experience

May 2025, Present

AI Researcher

York University, Toronto, ON
  • Designed and pre-trained transformer-based LLMs for tabular data from scratch, achieving state-of-the-art results on benchmark evaluation tasks.
  • Optimized multi-GPU HPC training with Hugging Face Accelerate and SLURM, cutting training time by over 60% versus single-GPU baselines.
  • Built end-to-end RAG systems integrating DeepSeek and Ollama with domain-specific vector databases for interpretable, context-aware literature querying.
Jan 2025, Apr 2025

AI Researcher

Kisoji Biotechnology, Waterloo, ON
  • Designed and fine-tuned a compact transformer for high-dimensional biological sequence modeling, cutting memory use by 30% while surpassing SOTA baselines.
  • Developed a prototype RAG-based code discovery system using vector retrieval, improving developer navigation efficiency.
  • Modernized a legacy codebase, replacing deprecated dependencies and restructuring modules to reduce technical debt.
Sep 2020, Oct 2024

Research Assistant

Western University, London, ON
  • Built large-scale biological sequence pipelines end-to-end, collection, curation, annotation, iterating on CNN and RNN architectures.
  • Designed sequence-based interaction models using BERT, T5, and custom Transformers, resulting in three peer-reviewed publications.
  • Engineered memory-efficient data generators for 700GB-scale datasets, reducing active memory use by 85% while maintaining throughput.
Jul 2019, Mar 2020

Machine Learning Engineer

Porsan Startup, Tehran, Iran
  • Applied Random Forest and XGBoost to large-scale survey data, generating insights for data-driven decisions.
  • Built a full-stack survey analytics platform with automated ML-driven feedback generation.

Research & publications

Self-Supervised Contextual Representation Learning for Transcriptomic Generative AI
A transformer-based self-supervised foundation model for high-dimensional transcriptomic data using masked token prediction and novel sample-specific tokenization.
ICML GenBio2026
Embedding Methods in Structured Data: A Comparative Study
A survey evaluating word, contextual, and protein embeddings across NLP and computational biology for protein interaction prediction.
Under review2026
TifBERT: Self-Supervised Representation Learning for High-Dimensional Data
A BERT-based model for self-supervised representation learning on high-dimensional structured data.
Under review2026
Component Puzzle Protein-Protein Interaction Prediction
An unsupervised puzzle-inspired framework using T5-based token representations for structural emulation, a 16% PR-curve improvement over SOTA.
Briefings in Bioinf.2025
Seq-InSite: Sequence Supersedes Structure for Interaction Site Prediction
A sequence-only model rivaling structure-based methods, up to 15% PR-curve improvement on interaction site prediction benchmarks.
Bioinformatics2023
PITHIA: Interaction Site Modeling with Attention and MSAs
An attention-based architecture incorporating MSAs, improving PR-curve by 35% and cutting memory use by 80% via generator-based data handling.
Intl. J. Mol. Sci.2022

More projects

Toolkit

Foundation models & deep learning

PyTorch JAX / Flax TensorFlow / Keras Hugging Face Transformers Transformer architectures Self-supervised learning Representation learning LoRA / PEFT

Generative AI & LLM systems

RAG pipelines LangChain Vector databases Embeddings Prompt engineering SFT / RLHF concepts DeepSeek Ollama

Structured & scientific AI

Sequence modeling World models Graph neural networks Scientific foundation models Protein representation learning Transcriptomic modeling High-dimensional data

Data & machine learning

Scikit-learn NumPy / Pandas PySpark SQL Data pipelines Large-scale preprocessing

Training infrastructure

Multi-GPU training Hugging Face Accelerate SLURM / HPC clusters Docker / Apptainer AWS Experiment tracking

Software & product engineering

Python C / C++ / Java React Flask Celery Redis Dexie.js Progressive Web Apps
The book

From Prompt to Response: an end-to-end journey inside LLM and RAG systems.

Sixty-one chapters tracing what happens between a typed prompt and a generated response, covering tokenization, embeddings, attention, context construction, retrieval augmentation, and modern LLM systems.

61 CHAPTERS · TECHNICAL DEEP-DIVE · LLM & RAG SYSTEMS