AI Engineer
Agentic Systems
Hello, I'm
AI Engineer & Automation Builder
I build AI systems that do real work — research agents that check their own sources, assistants that answer from your documents, and automations that handle routine business tasks without supervision.
Available worldwide
Islamabad, Pakistan
I'm Afnan. I build agentic AI systems from Islamabad — multi-agent graphs, retrieval pipelines over private documents, and automations that run business operations without anyone watching them.
Most of the work is the unglamorous half. My first research system produced confident, well-formatted reports that were quietly wrong. Fixing that meant scoring sources for credibility, checking each claim against the real text, and capping how many times it's allowed to retry. It now sometimes tells me it couldn't confirm anything — and that's the version I trust.
I also run a small AI studio called ZENTIX with two other engineers. Everything on this page I designed and built myself.
Education
BS Artificial Intelligence
COMSATS University Islamabad
Optimization
Builds a term's timetable from the courses, rooms, and teaching hours available. It works out where two classes would collide or a room would be double-booked, and rearranges around it instead of handing back a clash for someone to fix by hand.
View repository, opens in a new tab
Agentic AI
Several agents split a research question between them — one gathers sources, one drafts, one checks each claim back against the original text — and pass the work between each other until the report holds up. When a claim cannot be confirmed, it stays out.
View repository, opens in a new tab
Machine Learning
Estimates what a property is worth from its size, location, and features, learned from past sales in the same area. It also shows which of those factors moved the estimate up or down, so the number is not just asserted.
View repository, opens in a new tab
Computer Vision
Draw a digit and it reads it back. A convolutional network trained on handwriting samples, with its confidence shown next to the answer — so a messy 7 it is unsure about looks different from a clean one.
View repository, opens in a new tab
NLP
Reads product reviews and sorts them by what the customer actually felt, then groups the complaints together so the issue that keeps coming up is visible without reading every review.
View repository, opens in a new tab
Web Application
One place for running a residential society: who lives on which plot, what each household owes and has paid, and the complaints that are still open — so the office isn't reconciling a ledger, a register, and a stack of notes by hand.
View repository, opens in a new tabMulti-agent systems designed to research, reason and execute.
AI assistants that understand your files and surface what matters.
Connect your tools and eliminate repetitive manual work.
Purpose-built applications around your workflow.
I map how the work actually happens today, then find the parts worth automating.
System design first — what connects to what, and where the data moves — before any code.
Iterative, with working demos along the way. You see it running before it ships.
Live, with logging and error handling, so failures surface instead of hiding.
Every project starts with understanding the workflow — not the tech. The right architecture comes from the problem, not the trend.
Afnan
Let’s build something that actually ships
If there's a workflow someone is holding together by hand, or a pile of documents nobody has time to read — that's the shape of problem I'm good at. Tell me what you're building and I'll tell you what's realistic, including when the answer is that you don't need AI for it.
Start a conversation© Afnan Arif