AI Engineer · Karachi, Pakistan
I build AI agents that search documents, call real APIs, and answer the phone.
I work in Python and FastAPI, using the OpenAI Responses API, Agents SDK and Realtime API. I've built legal research bots, a multilingual mentor chatbot, WhatsApp and phone agents, and SaaS dashboards with a built-in assistant. Before LLMs, I trained CNNs to find tumors in X-rays and MRI scans.
Six builds that show the range: production client back ends, safe assistants, phone and voice agents, and messaging commerce.
Streaming chat backend for an AI mentor, deployed as three language instances. The provider (OpenAI, Grok, Gemini or Anthropic) switches at runtime without a restart. Answers are grounded in the mentor's books and video transcripts, and chats accept images.
Rolling summary every 4 user turns to cut token cost · CRM sync to GoHighLevel · GitHub Actions deploy to EC2 under PM2
A legal research agent over a curated store of judgement PDFs. The model runs its own multi-step file searches instead of one retrieve-then-answer pass, and falls back to labelled web search when a judgement isn't in the store.
Streams text.delta · tool.searching · tool.searched · done · every tool call and citation saved
A restaurant dashboard for orders, menu and table floor plan, with a chat assistant that answers questions about live data. The assistant is read-only by construction: no tools are sent to the model, so there is nothing to misuse.
Figures carry their period in the key, e.g. revenue_today_from_completed_orders_only
Call a normal phone number and talk to an OpenAI realtime agent. The server bridges Twilio Media Streams to a realtime session, lets callers interrupt mid-sentence, and tracks playback with Twilio mark events.
Call audio: G.711 μ-law, 8 kHz, buffered in 50 ms chunks
An LLM answer streams into Fish Audio's text-to-speech while it's still being written, so speech starts before the sentence ends. A second app takes food orders by push-to-talk, reads them back and places them.
Text is buffered 4 words, then 8, and never split mid-word
Customers pick a restaurant, browse a photo catalog, build a cart and check out inside WhatsApp. Staff use a dashboard to update orders, reply to customers and edit catalog products.
Native WhatsApp catalog orders · HTML confirmations over SMTP · replies in the customer's language
In my Calendar and Gmail agent, any tool that creates, updates, deletes or sends pauses for Approve or Deny. Reads go straight through.
If an assistant only needs to read, it gets no write path at all. A jailbroken prompt in RestaurantOS has nothing to call.
Users see tokens, tool searches and speech as they happen, over SSE or WebSockets, instead of waiting for a finished answer.
A wrong password and an unknown email return the same message, and someone else's session id returns the same 404 as a missing one.
Dawood University of Engineering & Technology, Karachi
Expected 2026
I'm open to AI engineering roles and freelance projects. Email is the fastest way to reach me.