The complete archive
Browse the full journey, or choose a layer to narrow the conversation.
I started by asking what Jev actually built. A four-subagent RAG example brought the opportunity into focus: useful decisions need useful context and software that knows what to do next.
By Kingsley Ijomah
The valuations attached to frontier-model companies will not rise forever. If they correct, the useful question for engineers is not whether AI disappears, but which parts of the stack remain worth understanding.
By Kingsley Ijomah
I keep hearing the same reaction around AI-heavy meetings: people leave tired, confused and unsure what was decided. The problem is not simply new terminology. It is what happens when teams discuss an AI solution before agreeing on the business problem.
By Kingsley Ijomah
I kept rebuilding the same agent inside different products. Stripping it back to one definition, three skills, one MCP tool and a visible loop showed me what an agent framework actually does.
By Kingsley Ijomah
I first read ‘Agents API’ as another agent-building SDK. The important change is that OpenAI now operates the Codex harness itself, including durable sessions, context compaction, recovery, sandboxes and subagents. Here is what that means for developers.
By Kingsley Ijomah
I started by comparing model providers, clouds, data platforms and agent frameworks as if they were fighting for the same prize. They are not. The more useful question is which control point each vendor wants to make hardest to replace.
By Kingsley Ijomah
I watched the confident advice swing from Claude to ChatGPT within weeks. It made me realise that the durable part of an AI workflow should be the context, process and checks you own, not whichever model happens to lead today.
By Kingsley Ijomah
When I first looked beneath a working RAG demo, I learned that retrieval quality depended on much more than a vector database. I built this local, inspectable project so you can follow the same path through chunks, embeddings, conflicts, deletion and evaluation.
By Kingsley Ijomah
AI can make every role faster while making delivery worse. The real opportunity is to turn specialist knowledge into reusable skills, redesign handovers and organise adoption around customer value.
By Kingsley Ijomah
I understood building with AI more clearly when a trend-detection project stopped being an LLM problem and became a system-design problem involving embeddings, time, evaluation, incomplete data and human judgement.
By Kingsley Ijomah
My first useful step into Building AI as a full-stack developer was not training a model. It was understanding what changes when software learns behaviour from data rather than following rules written by hand.
By Kingsley Ijomah
Everybody asks what they should use AI for. Almost nobody asks what they should not, because nobody selling it is incentivised to answer. The work it is worst at, why you cannot feel the difference, and what must never go in at all.
By Kingsley Ijomah
A better brief gets you a better draft. It does not tell you whether the draft is right. One in 277 peer-reviewed papers now cites work that was never written, and the checking nobody taught you is the part of the job that is left.
By Kingsley Ijomah
Everyone has a folder of prompts copied from somewhere, and it never works as well as they expect. The problem is the folder. Prompting is specification, not incantation, and four habits cover almost all of it.
By Kingsley Ijomah
Something it produced read well, went into a report, and turned out to be invented. Five behaviours explain why, and not one is a bug waiting to be fixed.
By Kingsley Ijomah
My brother, who has never written a line of code, and a software engineer with years behind him asked me the same question in the same month: I want to get into AI, where exactly do I start? There are three answers, they are different jobs, and the order you take them in matters.
By Kingsley Ijomah
