Research & Publications
I'm Ahmad Bilal, a Product Design Engineer & Researcher. I publish the research behind what I build, so anyone can check it. Three lines of work: whether an AI can stand in for a real person in user research, which the field calls synthetic user research; whether the test sets used to prove medical AI works are hiding a bias; and translation tools that turn English and Urdu speech into Pakistan Sign Language for deaf people in Pakistan.
Context-Free-Grammar (CFG)-based English/Urdu Translation System from Video & Audio to Pakistan Sign Language (PSL) Gestures
Deaf people in Pakistan cannot follow most video or audio. This system listens to English or Urdu and produces Pakistan Sign Language gestures. That is what Pakistan Sign Language translation AI does: it turns speech nobody in the room can hear into signs they can read.
Getting the words right is not enough. Sign language has its own grammar, so swapping word for word produces nonsense. So I wrote a context-free grammar for PSL, a set of rules every sentence has to pass before the system will sign it. The pipeline cleans up the audio, turns speech into text, then builds a sentence that satisfies those rules.
This is my only peer-reviewed paper. IEEE, ICICS 2024.
Grounded Simulation: Task-Preserving LLM Personas for Synthetic User Research
Grounded Simulation is my architecture for synthetic user research. It has an AI stand in for a research participant, but ties every answer back to real evidence and attaches a confidence score. A finding that cannot be traced to a source does not ship. I set it out in an SSRN working paper, written 1 April 2026 and posted on 28 April (doi:10.2139/ssrn.6503241).
One clarification, and it cuts against me, so it goes first. These two words have been used before. Two 2015 papers in the Journal of Artificial Societies and Social Simulation used “grounded simulation” for grounded theory combined with agent-based social simulation: Dilaver, who named it that, and Neumann, whose paper carries the title outright. A 2025 paper in the Journal for the Theory of Social Behaviour asks what grounded simulation is about human cognition. Robotics has grounded simulation learning, a third thing again.
None of them is this. What I introduced is the architecture, not the name, which is why the paper says “we introduce” rather than claiming the words. It is also not grounded theory, a method for analysing qualitative data, and not “grounded user simulation” from agent-evaluation papers in mid-2026, meaning a simulator tuned on logs of real behaviour. If you are citing any of this, cite the architecture and check the others yourself.
What is grounded simulation in UX research?
Grounded Simulation is a named architecture for LLM-based synthetic user research, introduced by Ahmad Bilal in an April 2026 SSRN working paper (doi:10.2139/ssrn.6503241). Its participants are built from behavioural science rather than invented: personality structure, the way different cultures answer questions, and how readily someone adopts something new. The interview never sees the hypothesis, so the AI cannot tell you what you hoped to hear. Every finding has to trace back to something a participant said. It is not grounded theory, which is a method for analysing qualitative data, and it is not the phrase “grounded user simulation” used in agent benchmarking.
The architecture behind Articos. The paper is under peer review, and now folds in my earlier prompting study “Less Expertise, More Coverage”. For the plain-language version, read the definition and mechanism in What Grounded Simulation is.
Auditing Class-Conditional Acquisition Confounding Across Five Open Tuberculosis Chest X-ray Corpora
Public chest X-ray datasets are how researchers prove a tuberculosis detector works. I audited the five most used ones and found the test can be won the wrong way.
The images come from different hospitals, and each hospital's equipment leaves a signature on the picture. A model can learn to recognise the hospital instead of the disease and still score well. To show that, I built a model that looks only at those signatures and never learns which patients had TB. It still scored 0.687.
Worse, the same test scores 0.990 or 0.569 depending on nothing but how you split the data. So the score often says more about the test than about the detector.
This is solo research, separate from my product work. Every label is radiographic and the paper makes no clinical claims. I released the audit tool, the notes on where each image came from, and the harder splits, so anyone can check a medical-imaging dataset before trusting it.