I'm a full-stack engineer working on the unglamorous half of AI products — the guardrails, the tenant isolation, the latency budgets, the human-in-the-loop approvals that stop an agent doing something expensive. I've shipped an autonomous task agent to the App Store, built a multi-tenant RAG platform end to end, and cut p99 response time by 64% on an orchestration engine I'm still tuning. I like the part where it has to actually work.
Shakeb — usually found shipping something.
About me
Most engineers pick a lane. I didn't — partly by temperament, mostly because building products alone doesn't let you.
I've refactored an enterprise booking platform's frontend until it loaded 30% faster, designed the Node APIs underneath it, and automated the pipeline that shipped it. Since then I've worked mostly on agents: an iOS task-execution agent that places real phone calls and sends real email on your behalf, and a multi-tenant RAG platform where the interesting problem was never retrieval quality — it was making certain one tenant's embeddings could never surface in another's results. The AI part is rarely the hard part. Making it trustworthy is.
“An agent that can act on your behalf is only as good as the moment it stops and asks first.”
Right now
Updated whenever reality changes — ask me about any of these.
A fast model answers while a stronger one reasons in parallel, then quietly corrects itself. Down to 436ms p99.
Several agents, one Next.js dashboard, and an ongoing argument with myself about how much autonomy is too much.
Dense pretraining, sparse upcycling, MoE. Mostly to understand what I've been building on top of.
Selected work
A few favorites — each one shipped, used, and still standing.
Making AI feel instant without making it dumber
A quick model starts answering immediately while a stronger one thinks in the background. When the better answer arrives it slips in as a correction instead of a jarring rewrite. You mostly don't notice it happen.
The agent that makes the phone call you're avoiding
You type one sentence — cancel the subscription, chase the refund, book the thing. It calls, waits on hold, and sends the follow-up email from your own address. It always asks before it acts.
Multi-tenant AI knowledge platform
Your email and your browsing become a searchable memory you can ask questions of. Built so that no organisation's knowledge can ever leak into another's answer — the part that's genuinely hard.
Readwise → Mem.ai, in one move
A small tool that moved thousands of saved highlights between two apps that didn't talk to each other. Both API keys stay in the browser session — nothing stored, nothing to leak. Mem's CEO emailed asking to make it a native integration.
Experience
Enterprise platforms, then a master's, then building alone. One constant: I ship.
What people say
“What sets Shakeb apart is his extremely composed approach towards his work, and his way of putting things out in the most transparent way possible … he is one of the most reliable person and delivers with exceptional quality. He has never backed down from anything where he can add value.
“He is the go-to resource for critical projects, seamlessly integrating into any team … He consistently delivers efficient and maintainable code, demonstrating professionalism, a strong work ethic, and innovative problem-solving abilities.
“Shakeb was always on top of his work, delivering always before time … With his intuition and creativity, he went over and above to add extra value to whatever was assigned to him.
Writing
I write to figure out what I actually think. It occasionally works.