Sulaiman Dauda

About

I build production systems with AI agents and infrastructure people run on their own servers, and I try to be honest about what both of them do.

I did not start in engineering. My degree is in business administration, and my first job title was project administrator. What changed things was building websites on the side from 2016 and discovering I cared less about how they looked than about why the third one was slow when the first two were not.

The middle of my career was spent as a data analyst in Lagos, at Sheffield Nigeria and then Bluechip Technologies. That is the part people skip over when they read the CV, and it is the part that shaped how I work. Analysis teaches you that the number you expected and the number you measured are different things, and that the gap is where the interesting problem is. I have never really stopped doing that, I just point it at servers now.

In 2023 I moved to the UK and took an MSc in Applied Data Science at the University of Essex. Since September 2024 I have led technical delivery on the platform at Sportsafe UK, a national B2B supplier selling to schools, clubs and councils.

Alongside that I build and open-source infrastructure in Go. Not tutorials or demos, but a hosting control panel that refuses a deployment which makes a site slower, a backend platform where PostgreSQL is the only infrastructure, a Docker Compose control plane whose central promise is that your containers keep running if it stops. All three exist because I wanted the thing and did not want to rent it.

Since June 2026 most of my code has been written with Claude Code and Codex. The work did not get smaller; it moved. Less of it is typing, and more of it is deciding what an agent is allowed to do, writing down the standards it works to, and checking what it produced against something that can fail. At Sportsafe that means a customer assistant whose hard rules live in code, an MCP server whose only write is a proposal a person approves, and evals that run in CI. The analyst habit turned out to be the right one for this too. The reply the model gave and the reply it should have given are different things, and the gap is where the work is.

I am in Essex, England, and open to senior and lead engineering roles where measured outcomes matter more than activity. The fastest way to reach me is email.

What I believe about building

Opinions I hold with evidence

Most infrastructure is added before it is needed

Across everything I have built there is no Redis, no message broker, no separate search service. PostgreSQL does job queues, full-text search, pub/sub and row-level security, and one database you understand deeply beats four you understand approximately. I would rather justify adding a dependency than justify keeping one.

A benchmark you cannot reproduce is marketing

A number is only worth reading if someone else can get it. So the hardware, the tuning applied to both sides, the exact commands and the raw output all go out with the result, including the metric where the competitor wins. If a comparison only ever appears as a bar chart, assume the method would not survive being shown.

Security is a boundary you can point at

“Secure” as an adjective means nothing. In Slipstream the unprivileged API cannot reach the system directly; it sends typed commands to a root agent over a socket, and commands are argv arrays rather than shell strings. In Windlass a lint rule fails the build if privileged imports appear outside one package. Boundaries that are checked automatically are the only ones that survive.

An AI agent’s output is a pull request, not a result

It gets read, tested and either merged or rejected, the same as a colleague’s. The tests that matter are the ones shown to fail first. A check that has never failed has never been tested, and the first one I wrote for a data-loss bug in my own tooling passed on the broken code.