AI Engineer — Rawalpindi, PK (UTC+5)
I build AI systems that hold up under real traffic. The hard part was never the model — it’s latency, edge cases, and what happens when something downstream fails.
as of Aug 2026
Answers every call, works out what the caller actually wants, checks real availability, and books it — then hands over to a person the moment it should. A missed call is a customer who already phoned someone else.
Answers pulled from your documents rather than the model’s memory — contracts, PDFs, meeting transcripts — with the source attached, so anyone can check it. Embeds in your product; your team controls what it knows.
Agents that do the work rather than describe it: they call your tools, write to your database, and stop when they should. Adding a capability means registering a tool, not rebuilding the system.
Reading about a voice agent proves nothing. Press the button and interrupt it.
01PKLatest project
Answers customer conversations, checks business opening hours, finds available appointment slots, books against Google Calendar with configurable slot duration, and notifies the team over Slack. It picks up on the phone through Twilio and on WhatsApp — the channels customers already use.
The hard part — latency. It worked, but took 1–2 seconds to start responding, which feels slow in a conversation. Four changes — a smaller low-latency model, cached availability, a stripped system prompt, and semantic caching — took it to roughly 300–400 ms.
Read the case studyFigures are from the current build and move as it ships.
Getting a language model to do something impressive takes an afternoon. Getting it to stay fast, stay cheap, and behave sensibly on the day something breaks is the actual work.
So before I call anything finished, I ask a different set of questions. How long does the user wait before they see a response? What does this cost per conversation at real volume? What happens when the calendar API times out halfway through a booking? Which of these calls genuinely needs a model, and which could be cached or handled without one?
One agent, three layers
generic example · not a specific project
Scroll the board sideways, then tap a node
Read the board
Three layers. The model lives on exactly one of them — everything above and below it is what decides whether the thing is fast, cheap and safe to put in front of a customer.
Hover any node — or tab through them — for what it does and why it is there.
Requests on the board
Those questions are where the engineering turns into business value. A receptionist that answers in 380 milliseconds feels like a person. The same system at two seconds feels broken — and the customer hangs up before it ever gets to be clever.
“I had received numerous proposals from freelancers before choosing Asad, and I have to say, most of them were either completely absurd or ridiculously overpriced. Asad was the only one who truly understood the project and provided a professional, well-thought-out solution. He was extremely available throughout the process, and we had regular calls to test and refine the code together. His approach was not only technically solid but also realistic and aligned with best practices. I highly recommend him to anyone looking for a skilled and reliable Python developer!”
“I have had numerous projects with Asad and I am satisfied each time… He has great communication skills. When he needs adjustment to the requirement, he would communicate and get confirmation.”
“Excellent engineer. World-class and top-grade. Very professional who puts efforts to ensure projects are excellent.”
“Asad showed exemplary technical capabilities, an excellent work ethic, and exceptional attention to detail.”
“Engineer is sound and has done many projects for us here at Upwork… Our rating is 5 STARS”
Full review history on Upwork
Nothing runs for three weeks and arrives as a surprise. Work goes back to build until you approve it, and you own everything at the end.
Scroll sideways for the full loop
Inputs, decisions, systems, failure points, approvals, and the exact action the system is meant to take — before any code.
If one agent with a few tools beats a multi-agent architecture, I’ll say so, and explain what the complex version would cost you.
Each independently testable, each ending with a short screen recording of what was built and how to operate it.
Your accounts, your keys, your numbers, from day one. Documentation for what your team can change without a developer.
8,760+hours in developmenta calendar year, to the hour
AI tools make the typing faster. They don’t get a vote on whether it ships.
I use them every day — scaffolding, test cases, reading a codebase I have never seen before. It would be strange not to.
What comes out still has to clear the same bar as anything typed by hand: a latency budget it actually meets, tests aimed at the edge cases rather than the happy path, and a sensible answer for the day a dependency fails. Fast on the parts that are typing. No shortcuts on the parts that are judgement.
Certification
Standing
Competition
Education
Language
Send the workflow you’re trying to fix, or the system that isn’t behaving. I’ll tell you the shortest reliable path to it — including when that path doesn’t involve me.
I’ll reply within an hour.