AI ENGINEER · AGENTIC SYSTEMS · MCP · RAG
I design agentic AI systems for the firmware and server-infrastructure world: MCP architectures, production RAG and LLM gateways. I write about AI, chips and supercomputers, and I mentor engineers trying to break into the field.
I started as an enterprise software engineer, came to the US for my master's in computer science, and landed in a corner of tech most AI people never see: the firmware and management layer that keeps servers alive.
That's where I learned to build AI that has to be right. My first project was the company's first RAG chatbot. It grew into a production agentic system, then a customer-facing product. Along the way I designed an MCP architecture that lets engineers plug AI agents into real enterprise and firmware APIs, and I demoed AI tooling on the floor at OCP Global Summit and Supercomputing.
Now I own an LLM gateway that routes production AI traffic across providers, and I spend my evenings writing and mentoring, because nobody handed me a map into this field and I'd like to draw one for the next person.
A Model Context Protocol store that lets engineers across the organization plug in modular AI agent integrations instead of writing one-off scripts, standardizing how enterprise IT and firmware APIs connect to LLM applications.
Architected and deployed a multi-step agentic RAG system over internal documents with zero external data uploads. Helped turn it into a customer-facing product, using RLHF feedback loops and LLM evals to improve retrieval accuracy and reduce hallucinations.
A LangGraph multi-agent app that lets engineers ask about CVEs and firmware/software package composition in plain English, built on vulnerability-management and SBOM APIs. Later converted into an MCP server.
Own an LLM gateway built on LiteLLM, now an external-facing SaaS platform: multi-provider support, observability and guardrails for production-scale AI traffic.
Built a conversational AI tool that reads and edits Grafana dashboards through backend APIs, demoed at OCP Global Summit 2023 and Supercomputing 2024, where it helped win a major server customer. Before that, built telemetry pipelines streaming Baseboard Management Controller logs into Grafana.
A Text-to-SQL app powered by an on-device LLM for a customer proof of concept: secure, low-latency querying of enterprise data with nothing leaving the machine.
Co-inventor on five U.S. patent applications, all pending. Five more inventions are in internal review.
Journal of Science & Technology. Early work applying ML and deep learning to medical diagnosis, and the start of my interest in AI for healthcare.
A new post every day starting November 16: MCP, shipping AI to production, firmware and infrastructure, chips, supercomputers, medtech, and notes for people starting out.
I mentor students and early-career engineers, especially international students and people moving from software into AI infrastructure. No fee. Bring your questions, your resume or your half-finished project.
contact@ai-khan.net
MCP. HTTP is a close second, and it knows it.
Read the logs. Then read them again. The answer is always in the logs.
Writing 100 posts in 100 days about AI, agents and chips.
Tool descriptions are the most underrated prompts in AI.
Agents, chips, supercomputers, and the secret second computer inside every server.
There's a working terminal at the top of this page. Try sudo hire mustafa.