AI/ML engineer at Invisible Technologies. I work on fine-tuning, preference optimization, and the data and evaluation work that decides whether a model improvement is real.
Website · Hugging Face · Kaggle · LinkedIn · Medium
I'm interested in how cheaply a model's behaviour can be changed, and how often the evidence that it changed is wrong. Four threads:
Weight-space editing. Whether closed-form, gradient-free edits to a checkpoint change what the model represents, and, separately, whether they change what it does. In llm-directed-weight-space-programming the representation geometry moves while the nearest-centroid decisions stay put. That gap is the part I want to understand properly.
Fine-tuning on a small budget. QLoRA runs sized for one free GPU, with benchmark leakage removed before training rather than confessed after. boxed drops the 7,345 GSM8K rows that ship inside NuminaMath-CoT, because otherwise the post-training score measures recall of the test set.
Tokenizer fairness. Urdu costs noticeably more tokens than English for the same content, which quietly shrinks the usable context and inflates the bill. I published a small token fairness report and the parallel corpus behind it.
Retrieval and access control. PolicyPilot treats authorization as part of retrieval instead of a filter applied afterwards, so the same question abstains for an employee and answers for HR. Most of the difficulty sits in the evaluation, not in the retrieval.
AgriMatch is a marketplace I built for my final year project: farmers list crops, buyers post requirements, and the two sides negotiate over WebSockets with a tracked order lifecycle. Next.js, FastAPI, PostgreSQL. It is the only project here that is a product rather than an experiment, and it taught me more about state machines than any paper did.
Occasional notes on experiments and evaluation on Medium.
arauf.bscsses21@iba-suk.edu.pk
Open to research collaborations and ML engineering roles, particularly on evaluation, interpretability, or multilingual work.

