Islamabad, Pakistan
I work on personalizing language models to individual users, at the memory level and at the weights level, and on how you evaluate a system that is supposed to keep changing. Currently building persistent per-user AI twins in production. Two IEEE publications on multi-agent deep RL for a hard constrained optimization problem, and a Mitacs Globalink term at McMaster University (2025) on integrating human advisory signals into PPO via subjective logic. Looking for a thesis-based MSc for 2027.
01 / Research
I am looking for a thesis-based MSc for 2027. My focus is LLM personalization: how a model adapts to one specific person over time, whether that adaptation belongs in memory or in weights, and how you evaluate it honestly once the target keeps moving. Reinforcement learning under realistic constraints, partial observability and noisy human feedback, is where I came from and still the toolkit I reach for.
Adapting a shared base model to individual users without training a separate model per person. Multi-LoRA composition and interference-aware merging, and hypernetworks that generate per-user adapters from a user representation instead of per-user fine-tuning.
Extracting durable conditional preferences from raw interaction history rather than storing transcripts. Association-rule induction with support and precision thresholds, embedding retrieval over past episodes, conflict resolution by specificity, and drift handling through validity intervals and decaying confidence.
How you measure a system whose target moves with the user. Rubric induction and rubric drift, LLM-as-judge bias, and the practical differences between DPO and group-relative policy optimization for preference learning at production cadence.
The methodological track underneath the above. Cooperative policy learning under partial observability, published in two IEEE venues, and human-guided policy optimization that fuses noisy advisory signals into PPO through a subjective-logic belief model.
02 / Background
I graduated from NUST in 2025 with a B.E. in software engineering and a 3.61 / 4.00 CGPA. My research track started in my third year at the Information Processing and Transmission Lab under Prof. Dr. Syed Ali Hassan, where I co-authored two IEEE papers on multi-agent deep RL for a hard constrained optimization problem.
In summer 2025 I was a fully funded Mitacs Globalink research intern at McMaster University with Dr. Istvan David at the SSM Lab. I worked on guided policy optimization: integrating human advisory signals into PPO via subjective logic belief modeling on Gymnasium PacMan, with measurable convergence gains over baseline PPO.
Since then my centre of gravity has moved to LLM personalization. I now work on persistent per-user AI twins at Alexein AI : assistants that carry a durable model of one person across months, and that have to be evaluated and improved without a fixed target to score against. Before that I spent a year at Adept Tech Solutions building production voice and email agents, which is where the retrieval, orchestration, and evaluation problems stopped being abstract.
I am pursuing the same question on two tracks. At the memory level: extracting durable preferences from raw interaction history and keeping them honest as they drift. At the weights level: multi-LoRA composition and hypernetworks that generate a per-user adapter rather than fine-tuning one model per person.
03 / Peer reviewed
Saleha Ahmed, Muhammad Uzair, Syed Asad Ullah, et al.
IEEE Internet of Things Journal · 2025
A cooperative multi agent DRL framework for CR-NOMA IoT, where distributed agents jointly learn spectrum access and power control policies under partial observability.
Saleha Ahmed, Muhammad Uzair, Syed Asad Ullah, et al.
IEEE International Conference on Communications (ICC) · 2025
DRL driven transmit power control for energy harvesting uplink nodes, evaluated against MRC, SC, and EGC diversity combining schemes under Rayleigh fading.
04 / Selected work
A REINFORCE policy-gradient agent on ALE Pacman with human advisory signals fused into the policy via subjective-logic belief modeling. Achieves measurable convergence speedup over baseline PPO.
Implementation backing two IEEE publications on cooperative MARL for wireless powered networks. From-scratch MADDPG, MATD3, MASAC, and MAPPO over a CR-NOMA environment with energy harvesting and three diversity-combining receiver variants (SC, MRC, EGC).
Proximal Policy Optimization on the MuJoCo inverted pendulum with a hand-shaped reward and Optuna-driven hyperparameter search. Groundwork for my Mitacs research on guided policy optimization.
Comparative study of single-device vs multi-GPU training for a learning-to-rank DNN on the ANTIQUE QA dataset, built on TensorFlow Ranking. Same ranking quality across both strategies, with a discussion of when MirroredStrategy actually buys you speed.
End-to-end IoT and ML pipeline built around a custom dataset gathered from ESP32 sensors deployed across campus. 75% AQI regression accuracy on held-out days.
Domain-specific LLM assistant for banking queries on LLaMA 3.2 3B Instruct with LoRA fine-tuning, FAISS retrieval, and live document ingestion. Testing how far a small open-weights model can go on a narrow domain when retrieval is well tuned.
05 / Timeline
Jun 2025 to Aug 2025
Hamilton, ON, Canada
Advised by Dr. Istvan David · SSM Lab
Jun 2024 to Sep 2025
Islamabad, PK
Advised by Prof. Dr. Syed Ali Hassan
Aug 2026 to Present
Islamabad, PK
Nov 2025 to Aug 2026
Islamabad, PK
06 / Academic
Nov 2021 to Jun 2025
Islamabad, PK
Relevant coursework
07 / Get in touch
available
Looking for a thesis-based MSc in machine learning for 2027. My research focus is LLM personalization: adapting models to individual users at the memory and weights level, and evaluating agents that are meant to keep changing. Reinforcement learning and learning from human feedback are the foundation underneath it.
If you are a professor or graduate admissions reviewer, email is the fastest way to reach me and I respond within a day. My CV is linked below.