About
Building toward AI sovereignty and an Open Frontier through domain-specific evaluations, agent safety, computer/mobile-use agents, and agentic robotics, including visuospatial memory.
Hey, my name is Kosi. I'm a technical founder and longtime hacker building at the seams between AI, robotics, and software.
My long-term work is centered on six bets: advancing AI sovereignty so individuals and organizations can control their models, data, compute, and deployment choices; building an Open Frontier through broadly accessible research, weights, and infrastructure; constructing domain-specific evaluations; improving runtime safety and security for AI agents; advancing computer- and mobile-use agents; and building agentic robotics systems, with a particular focus on visuospatial memory for physical AI. These are areas I'm committed to building and investigating, and this is where I document the engineering, research, and ideas shaping that work.
What I'm building
Pegasus
An open-source agentic runtime for long-horizon agents. It tackles the two problems I kept hitting in production: context compaction and reliability over long sessions. v0.1.7 is out, pip-installable, and defaults to Kimi K2 via OpenRouter. This is my long-term infrastructure bet.
Vidinie
An AI-powered video creation platform that turns written content into production-ready videos. FastAPI + Next.js + Remotion under the hood. Aimed at content writers who want their blog posts to live as video without becoming videographers.
Telemex
A fleet management and vehicle telemetry platform built through my KTP partnership at BCU and funded by Innovate UK. GPS map matching, OBD parsing, driver scoring, predictive maintenance, route optimization — the boring-but-critical infrastructure that keeps fleets honest.
How I work
I build from first principles. I'd rather understand a system end-to-end than wire together black boxes I can't reason about. This is probably why I keep ending up in places where the hardware, the firmware, the model, and the UI are all my problem.
I keep a personal multi-GPU cluster at home for experiments. I'm currently running K3s across two Ubuntu nodes with four RTX 4060 Tis between them partly because I wanted to learn it, partly because cloud GPU bills add up fast when you're iterating.
I think founders should ship things that work before they ship things that scale. Pegasus is the long bet; Vidinie is the near-term validation vehicle. Telemex pays the rent on my curiosity.
I take the ethics of building AI seriously, including the uncomfortable parts. If you're building agents that can do the work of multiple people, you owe it to yourself to sit with what that means rather than skipping past it.
Background
I did my undergrad in Computer Science at Babcock University, where I picked up the Most Promising Student Award on the way out. After that I worked on large-scale health data infrastructure at eHealth4everyone — Gates Foundation-funded work that integrated with the Nigerian Ministry of Health and the WHO. It was my first taste of building software that actually had to hold up under real-world weight.
I then did my MSc in Advanced Computer Science at Birmingham City University, graduating top of my cohort with Distinction. I was supervised by Dr. Moad Idrissi, and my thesis applied large language models to robot planning — work that's continued to shape how I think about agentic systems.
Along the way I've contributed to FastTrackML through the MLH Fellowship (working alongside G-Research engineers), led BCU's Formula Student Autonomous Driving team, won a hackathon for an AI-powered endoscopic surgery robot, and published research that's reached over 10,000 readers. I hold a UK Global Talent visa for contributions to technology.
I keep doing this kind of work because I'm an inventor at heart. Founder is the job; inventor is the disposition.
Contact me
If you'd like to talk about any of the topics I write about, or would just like to chat, you can reach out here: