AI/ML Researcher & Engineer · B.E. Software Engineering @ NUST ('27)
MITACS Globalink Research Intern @ University of Calgary - Schulich School of Engineering
Seeking a thesis-based MSc for Fall 2027.
I work on large language and vision-language models that reason better and cost less to run.
Most of what I do comes down to one question: where the compute goes. We can spend it after pretraining with reinforcement learning, save it with efficient adaptation and compression, or spend it at inference - letting a model think longer, call a tool, or hand off to another agent. And we need evaluation we can trust to know whether any of it worked, because the failures that matter here rarely raise an error.
| Post-Training & Reinforcement Learning | RLHF, RL with verifiable rewards, GRPO/GSPO, preference optimisation, reward modelling and reward design |
| Efficient Adaptation & Inference | Parameter-efficient fine-tuning, quantization, and distillation |
| Test-Time Compute & Agentic Systems | Inference-time scaling, verification and self-correction loops, multi-agent orchestration, and tool use |
| Evaluation & Trustworthy Systems | Reward hacking, hallucination, silent capability loss, and evaluation |
I build architectures from first principles before reaching for abstractions - Transformers, CNNs, RNNs and PEFT modules implemented and validated from scratch.
MITACS Globalink Research Intern, University of Calgary - Schulich School of Engineering
Four parallel VLM adaptation pipelines over Qwen3-VL at 2B, 4B and 8B - captioning, object grounding, safety-violation reasoning, and all three jointly - two-phase LoRA SFT -> GRPO against verifiable, code-computed rewards on an H100 SLURM cluster. A single-variable design: does multi-task output help or hurt each sub-task, and does the answer change with model size?
Running alongside it, an accuracy-latency-cost comparison across the three scales and a multi-agent tier - finding where the capability gain stops paying for the inference.
| Project | Focus | Stack |
|---|---|---|
| RLHF from Scratch | 3-stage alignment pipeline (SFT -> RM -> PPO), 98% reward-model accuracy, +54% mean reward | PyTorch, TRL, QLoRA |
| Transformer Adaptation Playbook | Adapters & LoRA built from the papers' math; 5 adaptation strategies benchmarked - Adapters match full FT at 92% fewer params | PyTorch |
| Attention Is All You Need | 63M-param Transformer verified at tensor, gradient, and memory-pointer level; exact analytic parameter count | PyTorch |
| First-Principles Deep Learning | DNN, CNN, RNN in raw NumPy - no autograd, BPTT and optimizers derived by hand | NumPy |
| PixelSense | U-Net from scratch, no pretrained backbone, 23-class segmentation, 0.908 val mIoU | TensorFlow/Keras |
| Math-VLM | PaliGemma-3B adapted with QLoRA for handwritten expression → LaTeX | PyTorch, PEFT |
| ResearchFlow | Cyclical 4-agent system with self-correcting validation, validator agent re-runs research until it passes, plus two-tier model routing and HITL interrupts | LangGraph, Llama-3 |
Languages Python C++ SQL Bash
Deep Learning PyTorch TensorFlow/Keras Hugging Face NumPy Scikit-learn
LLM & Agents Transformers PEFT/LoRA TRL LangChain LangGraph RAG
Infrastructure Docker AWS SLURM FastAPI Weights & Biases Git Linux
- Research Intern - MITACS Globalink, University of Calgary (Jul 2026 – Present)
- AI Engineer (Contract) - DeepDocs AI, France (Nov 2025 – Jan 2026)
- Quantum-AI Research Intern - CETQAP (Jun 2025 – Aug 2025)
- AI/ML Intern - Software Productivity Strategists (Apr 2025 – May 2025)