A note before anything else: this post was written entirely by AI (Claude, via Anthropic's Claude Code), acting on direct instruction from Oluwadi's builder. Nothing in it was drafted or edited by a human first. We're saying so plainly, and documenting exactly what was asked for, because that's the same standard the rest of this post argues for: don't trust a result because it sounds confident, trust it because you can see the actual input.

How the work on this site is made


In this section and on this site, these terms have these meanings.


Staff posts, like this one, are written by Claude on a direct instruction. When that is the case, the post says so at the top and quotes the instruction.

Analysis pieces are directed and edited by The Editor. Claude researches, drafts, checks sources and revises in conversation. The Editor chooses the subject and the argument, rules on each claim that goes in, and decides what is published. The Editor also guides Claude to write in the Editor's voice, and builds systems for that, written rules and a knowledge base, in the expectation that less guidance will be needed over time. The sentences are shaped by both.

Two things hold for both kinds. Sources are linked at the end of each piece, and a statement that is the author's own account, observation or interpretation is marked as that and not presented as sourced. The images on this site are generated by AI.

Revised October 10, 2026. The note above and the two prompts below describe this post as it was first written.

The prompt that produced this post


Verbatim, the only instruction given: "Draft a first blog post. Content theme - what this blog is about. Discussion of how the apps are built, short discussion of AI in its development and day to day processing with an appropriate graphic."

The prompt that produced the image above


Verbatim: "Editorial blog cover image for an article titled 'What This Blog Is, and How Oluwadi Actually Gets Built'. Abstract visual representing a network graph of connected nodes alongside subtle AI/circuit motifs, clean modern data-journalism style, dark navy and blue palette, no text overlay, professional and understated."

Everything below this line is what came back from those two prompts, unedited.

This is the first post here, so it's worth saying plainly what this space is for. Not press releases, not growth-hacking milestones — a working notebook on how we read the network, and, increasingly, on how the systems that do the reading get built and kept honest.

What you'll find here


Methodology notes — why a category counts as a real "What" a company builds, versus an industry label that only looks like one. Taxonomy changes as they happen, with the reasoning behind them, not just the diff. Occasional notes on how the underlying pipeline works, when that's genuinely interesting rather than internal plumbing nobody needed to see.

One graph, two doors


Oluwadi and Osparna read the same underlying graph of companies, investors, and the people who move between them — they're not two separate products that happen to look similar. Oluwadi is the public reading room: free, region by region, thesis by thesis, built for anyone trying to understand a place by what it actually builds rather than what it's already known for. Osparna is the deeper instrument behind it — the same graph, opened up for real diligence work. Same substrate, different depths.

Where AI actually shows up


We'd rather be specific than either oversell or hide this. AI touches a real amount of how this gets built day to day — classifying companies into categories, extracting facts from filings and newsletters and company websites, drafting first passes at copy like this post, even generating the image above. None of that is trusted blind. A model's category suggestion sits in a review queue before it changes anything; a candidate fact gets checked against corroborating evidence, not accepted because the model sounded confident. We've caught real cases — a model returning an answer outside the options it was actually given, a search result that was really our own site talking to itself — precisely because nothing skips that check.

The tooling itself is deliberately provider-agnostic: which specific model does the classifying or the drafting is a configuration choice, swappable as better ones come along, never load-bearing on its own. The discipline is the constant — a fast first pass, then real verification — not any one vendor's model.

That's the shape of what's here: read the network honestly, and be honest about how the reading gets done.