Collins Madubuilds data pipelinesweb platforms

€70 an hour as a baseline, and that covers the whole job — architecture, backend, frontend, deployment and the documentation someone else can run it from. I don't bill those as separate rates. Where a project can be scoped properly up front I would rather quote a fixed price, because that puts the risk of my own underestimate on me instead of on you.

Usually within a week, sometimes the next day. The variable is rarely availability — it is how much research a project needs before I can say anything useful about it. A landing page I can scope in an afternoon; a system that has to read supplier spreadsheets in nine languages takes rather longer before I would be willing to quote it.

Most run four to twelve weeks. The two largest have been going for over a year — once a system is something people depend on every morning, there is usually a next thing worth building.

Production systems for businesses: procurement and pricing platforms, approval and e-signature workflows, employee portals, AI discovery products, scraping and ETL pipelines, and the dashboards that sit on top of them. Web, mobile, and backend — usually all three on the same project.

Regularly. Several projects here started as someone else's repository — a web frontend I extended into a full mobile product, an e-commerce build I picked up mid-flight, a marketplace scaffold I forked and rebuilt around a new ML service. Reading unfamiliar code is most of the job.

Mostly solo. Almost everything in the archive I designed, built, deployed and documented alone — including projects I inherited part-way and owned from that point on. A few were shared builds where I took the front end and someone else the back, or the reverse; those carry the role that was actually mine and claim nothing beyond it.

They're client systems. Those entries describe the problem, my contribution and the stack, and identify the client by sector rather than name — a European B2B cosmetics distributor, a capital-investment firm. No source, no screenshots, no data. Happy to go deeper in a conversation.

All of it runs. Fine-tuned segmentation models with a tiled-inference scheme, vector search over Milvus, LLM extraction with schema validation and fail-closed handling, a vision pipeline sustaining 300+ listings a minute. Where a model is doing something a rule could do better, I use the rule.

A deployed system, the repository, and documentation written for whoever runs it next: architecture notes, a runbook, and operator SOPs. On one internal product that included a full documentation site with annotated walkthroughs.