Final year, BS Artificial Intelligence at UMT.
I take a build from an empty repo to a live URL, then write down whatever broke on the way.
I spent a week reading retry code. One provider account had gone dead and every job in the queue was waiting behind it. The fix was nine lines. The week was not.
Lahore is home. I am in the final year of a BS in Artificial Intelligence at the University of Management and Technology, expected 2027, and most of what I know arrived through shipping rather than coursework. Freelance work on Fiverr started in February 2023 and ran to March 2025: five plus web apps for clients in the USA and Europe, 100% completion, a 5 star rating. One of them was a console resale marketplace with no payment rail at all. A seller submits the device details, the platform quotes, and once the quote is accepted both sides carry on over email.
Mailtrail was a contract, December 2024 to January 2026. An organization adds its employees and students, an LLM writes the test, the test goes out from that organization’s own email domain, and graded answers come back into React dashboards. Redis backed queues carry dispatch and evaluation. It runs on AWS, where rewriting the slowest database queries took about 30% off the monthly bill. In the same stretch of freelance work I took a legacy Node.js and MongoDB e-commerce backend from 3.2s to 400ms with indexing, query work and Redis caching, and moved another client off managed cloud onto Docker for 55% less hosting.
Kairo was built at Grow_in between March and July 2026. React and TypeScript in a Tauri desktop app, FastAPI underneath, and a pipeline in three stages: a Writer producing 10 scene scripts through OmniRoute LLM routing, a TTS module aligning narration with edge_tts word boundary timestamps, and an Aligner checking each scene against a 2.25 words per second baseline. EDLs generate from the JSON scene definitions, so nobody assembles a timeline by hand. Manual production effort fell by 80%. I also built Fable, a Shopify app with three custom components, and put the stack in Docker with CI/CD.
Cetus is mine. Flow, Qwen and edge-tts sit behind one internal API, so image, video and voice generation all run from a single composer. FastAPI and async SQLAlchemy 2.0 on Postgres through Supabase, Alembic migrations, an in-process asyncio worker pool, mypy strict over the provider, auth and credit code. The frontend is Next.js App Router with TanStack Query, and one Server-Sent Events stream carries the progress of every job. Auth is Supabase with Google OAuth, billing runs on credits, and an admin console and usage analytics sit behind it. Web on Vercel, API on GCP, Postgres in ap-southeast-1. It is live at cetus-one.vercel.app.
CyberBrain IDS is the one that never had users, only traffic. A Random Forest and XGBoost ensemble reads live network traffic over a 5 second rolling window across 22 plus packet features. A RAG pipeline on an LLM and Chroma writes the threat report and maps it to MITRE ATT&CK, and a React SOC dashboard streams incidents over WebSocket with six turns of follow-up questions available per incident.
The research is one paper, first author, accepted at ICATCICT 2026 in Dubai. It is set out further down this page. The blog is the other half of the record: 5 war stories so far, each one taken from a real commit, including the week lost to a dead provider account and the payments page that shipped bank details for an account that never existed.
BS Artificial Intelligence
Final year. Expected 2027.
Five plus web apps, 100% completion, 5 star rating.
Tests written by an LLM, sent from the organization's own domain, graded back into React dashboards.
Three stage pipeline and automated EDL generation. Cut manual production effort by 80%.
Flow, Qwen and edge-tts behind one API, with credits, auth and an admin console on top.
Accepted, to appear in AIP Conference Proceedings.
Accepted at ICATCICT 2026, Dubai, UAE, 25 to 27 November 2026. To appear in AIP Conference Proceedings, indexed by Scopus and Web of Science.
