When someone asks the Mave Agent about their market, the desktop app sends the question to a planning model, which picks tools exposed over MCP. Retrieval pulls passages from the client's documents and meetings, and the answer streams back with its sources. The whole path runs on Kubernetes, which I operated as the platform grew.
who is sheharyar?
I AI and the systems they on.
Hover an underlined word, or
Lead engineer at Mavera for two years, on an AI platform used by 120+ companies. Now in the final year of an MSc in Computer Science at Sorbonne, Paris.

Software I led is used by teams at
In numbers
- $100K+Monthly revenueCarried by the platform I led at Mavera.
- 165Client projectsShipped as a freelancer, rated 4.9.
- 1stMistral hackathonLLM × Law, Paris, October 2026.
- 1/17System DesignTop of the class at Sorbonne.
What each word is backed by
My one-line summary, read the way a language model reads it. Hover a word to see what it pays attention to.
pulls on everything after it, the way a subject should.
- Lead engineer at Mavera Two years, 120+ companies0.31
- MSc at Sorbonne Final year, Paris0.22
- First place, Mistral hackathon Team of five, October 20260.19
Selected work
Bina.ai
Mistral LLM × Law Hackathon, Paris, October 2026
First place, team of fiveSee the project: Bina.aiHosted by Mistral AI, sponsored by Stanford Law School, Legora, Hector AI, Jus Mundi and Sciences Po.
Bina turns French data-protection rulings into briefs for litigation funders. A funder can't act on a brief that misquotes the ruling, so the backend I built checks every quoted sentence, word for word, against the original, and keeps any it can't find out of the pitch.
The regulator found that “the company did not implement measures appropriate to the risk” and noted that “customers were never informed of the breach”.
Press the button to check both quotes.
… the restricted committee considers that the company did not implement measures appropriate to the risk, within the meaning of Article 32 of the GDPR. The customers concerned were informed by email on …
Someone describes an AI tool in plain words. An agent builds it as a working web app on the customer's own domain, and every app runs on one platform that routes its requests to 38 models from five providers, on the owner's own keys.
What people I've worked with say
He moves quickly, asks sharp questions, and has a strong ability to take early ideas or partial direction and turn them into well-structured, working systems.
Sheharyar is the kind of person you can trust with the hard stuff. He's thoughtful, dependable, and quietly makes everyone around him better.
Everything I've built, on one map
Placed by what each project is about, the way an embedding model would. Each small dot is one of 165 freelance projects.
Mavera
AI market-intelligence platform. I built it end to end, then led the developers who joined.
120+companies using it
- Next.js
- Rust
- MCP
- Kubernetes
Now
Sorbonne Université, 2025 to 2027
This year I'm going below the agents: GPU programming in CUDA, parallel computing with MPI, and deep learning for vision. Last year I ranked first of seventeen in System Design and validated a Rust traceroute engine at the LIP6 lab.
March 20255 min read
85Kviews on LinkedInJuly 20253 min read
March 20252 min read

I like knowing what happens under the button.
I started freelancing in the second year of my degree and shipped 165 projects for clients. Then came an operations platform for a restaurant group in Dubai, a document editor for researchers in Finland, and two years leading engineering at Mavera, all remotely, before I moved to Paris for my master's.
- Based in
- Paris
- Languages
- English, French (learning)
- Strengths
- LLM agents, retrieval, Rust, Kubernetes
- Learning
- CUDA, MPI, PyTorch
- Available
- From 15 February 2027
- Paperwork
- Convention de stage
Let's build something together.
Six-month internship in Paris from 15 February 2027. Sorbonne signs the convention de stage, so there's no visa paperwork on your side. I reply within a day.