Open to PhD — 2026/27 intake Computer Vision Generative AI Efficient ML

PhD Application
Irfana Aslam

MS Computer Science · Peer-reviewed publication (Impact Factor 4.5) · MS thesis at above 98% accuracy · 40 shipped engineering projects · 450+ hours of teaching. This page is written for prospective supervisors: everything needed for an initial assessment is here, without requesting documents first.

1Peer-reviewed publication
IF 4.5Journal impact factor
above 98%Thesis model accuracy
40Engineering projects
450+Lessons taught
C1English proficiency
Research Statement

Learning from data you cannot collect

My research has circled one problem from two very different directions, and I did not notice the pattern until recently.

In my MS thesis on fabric defect detection, the obstacle was never the network. Reviewing two decades of prior work — auto-encoders, transfer learning, regional CNNs, U-Net, attention mechanisms — showed strong reported accuracies that each collapsed to a narrow set of defect classes or a single fabric pattern. The binding constraint was the dataset: defects are rare by definition, so real collections are small, imbalanced and inconsistently labelled. My response was to stop treating the data as fixed. I built a generator that synthesises defects into pristine fabric with controlled variation in type, size, shape, orientation and lighting, producing 100,000 exactly-labelled images across ten classes. The resulting model reached above 98% accuracy, beating SVM, Random Forest and a baseline CNN on the same task.

Independently, I later built BitWithBite AVS — a virtual production engine that turns a written prompt into a rendered film. It exists because I wanted to make animated educational content and could not afford a studio. Only after it worked did I recognise the same instinct at work: where the data does not exist, generate it under controlled conditions. AVS synthesises entire scenes rather than defect patches, but the epistemics are identical — and so are the open questions.

Those open questions are what I want to spend a PhD on. Synthetic data always carries a domain gap, and my thesis is honest about not having measured its own rigorously. When a model is trained on generated examples, what exactly has it learned — the phenomenon, or the generator? How do we measure that gap, and can generative models close it rather than widen it? These questions matter most precisely where data is scarcest: industrial inspection, medical imaging, rare-event detection — domains where the examples that matter are the ones you have least of.

I bring an unusual combination to this: research training and shipping discipline. I have written a peer-reviewed paper and a thesis, and I have also built and maintained 40 working systems, several running in production at zero marginal cost on commodity hardware. I care about work that runs, and about being straightforward regarding what it does not yet do.

Research Interests

Where I want to go deeper

🧪

Generative Data Synthesis

Producing training data for rare events. Measuring and closing the synthetic-to-real domain gap; comparing parametric injection against diffusion-based generation.

👁️

Computer Vision & Anomaly Detection

Reference-free anomaly detection that generalises past known patterns — reconstructing expected appearance and flagging deviation, rather than classifying a fixed defect taxonomy.

🎬

Generative 3D & Automated Cinematography

LLM-driven scene synthesis: turning natural language into structured 3D scene representations, camera language and lighting decisions that hold up as coherent visual narrative.

🏗️

Generative Design with Performance Feedback

Closing the loop between form generation and physical simulation: coupling single-view 3D reconstruction with building energy modelling so that a designer's sketch is evaluated, not just rendered — and the trade-off is presented as a Pareto front rather than a single score.

🤖

Multi-Agent LLM Systems

Composing specialised models — writer, designer, critic — into pipelines where intermediate state stays inspectable and human-correctable rather than opaque.

💚

Efficient & Low-Resource AI

Capable systems on commodity CPU hardware without paid API dependencies. A research equity question as much as an engineering one — who gets to do this work.

🧬

AI for Scientific Discovery

Building on my immunoinformatics publication: computational methods that reduce expensive wet-lab search spaces before physical experimentation begins.

Proposed Directions

Questions I would like to work on

These are directions rather than a fixed proposal — I would expect them to change substantially in conversation with a supervisor whose programme they intersect.

Direction 1

What does a model trained on synthetic data actually learn?

When defects are injected parametrically, a classifier may learn the injection process rather than the physical phenomenon. I would like to develop diagnostics that separate the two — measuring how much of reported accuracy survives transfer to real data, and identifying which generator parameters drive genuine generalisation versus shortcut learning. My own thesis is a natural first case study, since I can rerun it honestly.

Direction 2

Can reference-free anomaly detection replace fixed defect taxonomies?

My thesis assumed a known reference pattern, which works for repeating printed fabric and fails on irregular weaves or fabric under tension. A reconstruct-and-compare formulation would generalise, but raises its own question: how do you set a deviation threshold when you have never seen the anomaly class? In industrial settings recall on defects matters far more than accuracy — a false alarm costs a second look, a miss costs a roll.

Direction 3

How do you evaluate a generative pipeline with no ground truth?

AVS produces films from prompts. There is no correct output to compare against, which makes evaluation genuinely hard — and this is the shared problem across generative systems. I am interested in evaluation frameworks for open-ended generation, including whether a critic model can substitute for human judgement and where that substitution breaks down.

Direction 4

What is the real accuracy cost of running AI on constrained hardware?

I build for CPU-only machines by necessity, which has made the trade-offs concrete. I would like to characterise them properly: quantisation, distillation and architectural choices measured not on benchmark accuracy alone but on task-level outcomes under a fixed compute budget — the constraint most of the world actually operates under.

Evidence

What I can already demonstrate

Peer-reviewed publicationMicroorganisms (MDPI), IF 4.5 — experimentation, data generation, manuscript drafting and revision
Completed MS thesisabove 98% accuracy, benchmarked against three baselines, full figures and per-class metrics published
Independent research capabilityAVS was conceived, designed and built alone — 39 modules, ~9,900 lines, 29 films produced
Research-grade engineering practiceAI-CAD, a 54-module sketch-to-energy design pipeline — reproducible provenance, a 28-check test suite, and a quantitative evaluation suite (Precision@K/mAP, Chamfer/IoU, method-agreement statistics)
Programming depthPython, TensorFlow, OpenCV, FastAPI, Blender bpy — production systems, not notebook experiments
Teaching experience450+ lessons, 445+ hours, 59 students — directly relevant to TA and demonstrator duties
Scientific writingThesis authored in full; contributed to manuscript drafting and revision on the published paper
Cross-disciplinary backgroundBS Bioinformatics into MS Computer Science — comfortable moving between domains
English proficiencyC1 across listening, reading, writing and speaking

Research trajectory

2017 – 2021
BS Bioinformatics — computational biology foundations
Sequence analysis, genomics, BioPython. Groundwork for the immunoinformatics publication.
2023
Published — multi-epitope universal coronavirus vaccine design
Microorganisms (MDPI), Impact Factor 4.5. Contributed experimentation, data generation and manuscript writing.
2021 – 2023
MS Thesis — fabric defect detection with deep learning
Synthetic dataset generation, custom CNN, above 98% accuracy benchmarked against SVM, Random Forest and baseline CNN.
2024 – 2025
Applied LLM & RAG systems
Production retrieval-augmented assistants with FAISS and local models — grounded answers with citations at zero API cost.
2025 – Present
BitWithBite AVS — generative 3D and automated cinematography
Multi-engine pipeline converting natural language into rendered film. Ongoing work on low-resource inference and scene synthesis.
Documents

Everything, up front

No need to request these separately.

Working Together

What I would bring to a research group

🔨

I build things that run

40 shipped projects, several in production. Ideas become working systems rather than staying in slide decks — useful for a group that needs prototypes, tooling or infrastructure to actually exist.

🧭

I work independently

AVS was conceived and built alone across six development phases with no external direction. I can be given a hard problem and a long horizon.

📣

I can teach

450+ hours of one-to-one instruction across every level. TA duties, demonstrating and supervising undergraduate projects are already familiar territory.

🎯

I am candid about limitations

My own case studies document what my work does not yet do and what I would change. I would rather report a real limitation than defend a number.

🌍

I work well remotely

5+ years with teams in Canada and the United States across time zones — asynchronous collaboration is my default mode, not an adjustment.

💰

Funding

Actively seeking funded positions, scholarships and assistantships. Previously awarded PEEF, EHSAAS and eHunar scholarships on academic merit.

Let's talk

If any of these directions overlap with your group's work, I would welcome a conversation — and I am glad to adapt them substantially toward an existing programme.

irfanaaslam@bitwithbite.com  ·  +92 329 910 6379  ·  Lahore, Pakistan