I build applied AI systems end-to-end — from data and research design through to models, APIs, and deployed products. My work spans LLM/RAG pipelines, applied ML for signal classification and clinical prediction, and the full-stack interfaces that make model outputs actually usable.
I'm strongest where research has to become software: turning experiments into reproducible pipelines, exposing models through practical FastAPI apps, and documenting projects so another engineer can run, inspect, and extend them.
|
LLM & Applied ML RAG pipelines, prompt engineering, cited answers, classification, regularisation & generalisation studies, imbalanced learning, formal evaluation. |
Product Engineering Python, React, FastAPI, Streamlit, REST APIs, auth, dashboards, and user-facing AI workflows deployed on Azure & Vercel. |
Engineering Discipline Clear READMEs, reproducible setup, honest metrics, safe secret handling, and pragmatic, deployable architecture. |
Every project below is live — click a demo to try it.
|
Turn any website into a private, source-backed AI expert. Crawl a site → index it as vectors → chat with answers that cite their sources. End-to-end RAG behind a clean FastAPI service.
|
Reproducible ML study across real EEG datasets. How preprocessing order & regularisation jointly affect generalisation on imbalanced EEG — 3 benchmarks, 5 hypotheses tested for statistical significance. Code →
|
|
Media-literacy platform for spotting AI-generated content. Interactive learning around C2PA metadata, provenance chains, and deepfake detection — making AI-detection skills accessible.
|
Modern cybercrime-reporting cell for Pakistan. (collab) React + FastAPI prototype for citizens to report incidents and track cases, with local ML tools for flagging AI-generated media.
|
- Under review (2026) — Isolating an interpretable, length- & code-controlled linguistic signal in LLM prompt-injection and jailbreak corpora. (methodology to follow post-publication)
- Applied study — Chronic Kidney Disease risk prediction with a leakage-safe scikit-learn pipeline on the UCI dataset, served via FastAPI.
| Area | What I care about |
|---|---|
| Reproducibility | Documented data sources, dependency files, and experiments another engineer can rerun from scratch. |
| Model credibility | Honest metrics, validation boundaries, baseline comparisons, and clearly stated failure modes. |
| Product usability | Interfaces that expose model behavior clearly instead of hiding it inside notebooks. |
| Repository quality | Clean READMEs, safe secret handling, sensible structure, and focused, deployable projects. |
- Shipping deployable AI products (LLM/RAG, media forensics) rather than notebook-only experiments.
- Publishing my LLM prompt-injection research and strengthening reproducibility across repos.
- Building honest, well-documented ML — clear metrics, validation boundaries, and clean setup.
- Open to freelance AI/ML work, software-engineering roles, and research collaboration.
