Available for full-time roles and project work
experiences.
3+ years of industry experience across machine learning, deep learning and software engineering — from data collection and model development through to the applications people actually use.
Neural core · generated in real time
About
I spent 3+ years as an AI Engineer at AI4LYF working across the full machine-learning lifecycle — preparing data, running experiments, developing models, evaluating whether the results could be trusted, and integrating the outcome into software people could operate.
That work spanned deep learning and classical machine learning: a Pix2Pix GAN for image reconstruction, CNN classifiers trained on hybrid real and synthetic datasets, gradient-boosted models for biomedical audio, and document automation built on computer vision and OCR. It also meant designing the data collection framework behind a 20K+ sample dataset, and the backend and validation workflows that kept it usable.
Since October 2024 I have been working independently — building and deploying application prototypes across real estate, education, healthcare, retail, hospitality and packaging. The six projects in this portfolio come from that period. They are where the software engineering, the product thinking and the interface work meet.
A model that nobody can use is not a finished piece of work. The interesting engineering is usually in the distance between a result and something someone can operate.
BS Software Engineering, The University of Lahore · 2014 — 2019
AI Engineering
Three and a half years at AI4LYF, working the whole lifecycle rather than one slice of it. Select a stage to see what that stage involved.
Stage 01 · Research
Understanding what a domain actually needs before choosing an approach — reading the problem, the constraints and what the available data can honestly support.
Generative Deep Learning
A GAN-based image reconstruction system for imaging enhancement, built and optimised across MAE, VGG, content and classifier-based loss functions to find the objective that produced the most usable reconstruction.
Computer Vision
Multiclass and multilabel classification models trained on hybrid datasets that combined real samples with GAN-generated ones, alongside an assessment of how far synthetic training data could be relied on.
Automation · OCR
Automated document digitisation and data extraction, combining classical image processing with OCR and a managed extraction service to cut manual data-entry work out of the loop.
Applied Machine Learning
Gradient-boosted and ensemble models for classifying biomedical audio across four clinical conditions, where engineered features and tree ensembles outperformed the reach for a deeper network.
ML Infrastructure
Designed and led a framework that produced 20K+ validated samples — the backend application behind the collection, the validation workflow that kept it clean, and the training that kept the team consistent.
Data Applications
Interactive dashboards in Python Dash that put model output and biomarker data in front of the people who needed to act on it, improving reporting and visual analysis workflows.
Generative AI
The first strips RAG down to visible code — no framework, every stage written out. The second rebuilds the same problem with the retrieval, conversation handling and failure behaviour a real deployment needs. Both are live, and both are open source.
No framework — every stage written out by hand
Point it at a URL or drop in a PDF and it builds a searchable index in the browser session, then answers strictly from what it retrieved. Written without a RAG framework so each stage — cleaning, chunking, embedding, ranking, grounding — is visible code rather than a default.
Pipeline
Engineering decisions
TF-IDF over scikit-learn and dense embeddings via sentence-transformers (all-MiniLM-L6-v2), both scored by cosine similarity. Having the sparse baseline next to the dense one is what makes the difference between them measurable rather than assumed.
The same problem, engineered for production
A conversational chatbot that answers only from an indexed corpus and cites the passage behind every claim, down to the page. Built as an LCEL chain on LangChain 1.x with Gemini and a persisted Chroma store, split into an offline ingestion pipeline and an online query pipeline.
Pipeline
Engineering decisions
“What about the second one?” carries no signal for a vector search. A cheap model condenses the turn against the history into a standalone query first — and a RunnableBranch skips that round trip entirely on the first turn, where there is no history to resolve.
Selected Work
Six applications built and shipped independently. Each one is live, each one is open source, and every feature described is in the running build.
The Pack Style · Concept build
A packaging and print studio site built around one continuous 3D world. A single procedural rigid box carries the entire page — the camera, the lid hinge and the dieline unfold all interpolate from the same scroll journey.
Smart Cash & Carry · Interactive demo
A grocery storefront where search and category filtering swap the entire page between a curated home layout and a live results view — instantly, with no transition to wait out.
Ghazali Public School · Website & portal prototype
A public school site with three role-specific portals behind it. Parent, student and teacher each get a different view of the same record, and the homepage preview of each portal is drawn with the real dashboard components.
Faisalabad Property Discovery · Platform concept
A property discovery concept fronted by a procedurally generated 3D skyline. Residential and commercial filtering, project showcases, a locations explorer, payment plans and lead capture.
Technical Stack
Grouped by the job they do rather than listed as a wall of logos. Open a group to see the full set.
Classification, regression and evaluation work, from feature engineering through to judging whether a result can be trusted.
What I Build
Machine-learning and deep-learning development end to end — data preparation, experimentation, model development, evaluation and the integration work that turns a model into something usable.
Applications built on language models: retrieval-augmented generation, prompt design and the surrounding software that makes an LLM feature dependable rather than a demo.
Python backend and application development — APIs, database integration and the services that sit between a model and the people using it.
Typed React and Next.js product experiences, responsive from phone to desktop, deployed continuously — the discipline behind every project in the work section.
Document digitisation and workflow automation using computer vision, OCR and cloud extraction services, aimed at removing manual data entry from a process.
Real-time WebGL built into a product rather than bolted on — procedural geometry, custom shaders and scroll-directed scenes that stay fast on a mid-range phone.
How I Work
Get clear on the problem and what the business or product actually needs, including what the available data or constraints will realistically support.
Define the architecture, the interface and the interaction strategy — decide the shape of the thing before writing the parts of it that are expensive to change.
Implement the system: models, services, application code and the interface, in working increments rather than one large reveal.
Deploy it, test it against real use, then optimise and improve — performance, accessibility and the details that only surface once something is live.
Experience
Three and a half years of professional AI engineering, an independent development period, and the internships that preceded them.
Education
The University of Lahore
Continuous learning · AWS
Learning plans and courses, not certifications.
Additional Work
Different industries, the same engineering discipline — each one solving a problem the category usually skips.
Moon Grill · Interactive demo
A two-branch restaurant site where the chosen location persists and shapes the rest of the visit — a large menu with search and virtualised rendering, a basket, reservations and a gallery.
Vessante Aesthetic Institute · Concept build
A clinic concept where the imagery is synthesised rather than sourced — a GLSL silk hero, SVG-generated tissue surfaces behind the before/after slider, and a lathed glass bottle you can spin.
Contact
Let’s turn it into a high-quality digital product. Tell me what you have in mind and I’ll come back with how I’d approach it.
Based in Faisalabad, Pakistan · Available for full-time roles and project work