Veridian: turning 4M words into an A* on my own exam

Veridian started as a selfish project, and I don't think it would have worked if it hadn't. I wanted an A* in Pearson IAL Economics & Business, and every past-paper resource I could find fell apart at exactly the point where it mattered most: it could tell you an answer was wrong, but not why an examiner would dock it a mark, not what the grade boundary actually demanded, not how the assessment objectives were meant to be weighed against each other. Marking is a discipline with its own hidden rules. Nothing available taught them.

So I built the thing I actually needed. Not a summary of the syllabus — a knowledge base deep enough to mark the way an examiner marks: 4M+ words across 200+ documents, 40 to 65 per paper, covering the assessment objectives, the grade boundaries, the UMS conversions, all of it. Then I wired that knowledge base to an AI marking harness that could take a full paper and return the kind of feedback an examiner gives — a grade, an AO breakdown, and a confidence flag on the calls that were genuinely borderline, instead of a single number with no explanation behind it.

Building it was slower than studying would have been on its own, and there was a long stretch where it wasn't obvious the time would pay back — a knowledge base that size doesn't get useful until it's most of the way finished, which meant most of the work happened before I had any proof it would work at all. There's a specific kind of doubt that comes with that: you're not failing at the thing you set out to do, you're just not done yet, and from the outside those look identical for far longer than is comfortable. The only way through it was treating the unfinished tool as if it already worked, and fixing what broke when it didn't — which is a different discipline from studying, where the feedback loop is fast and someone else already built the test.

It got me a 98% and the A* I was after. That part, on its own, would have been enough to call it a win and move on.

The second lesson only showed up after the grade, and it's the one I keep applying since: once a tool works that well for you, the need behind it was never really only yours. I hadn't built Veridian to sell. But the gap I'd found — good marking-grade feedback simply not existing for this board — wasn't a personal problem, it was a market one, and I was sitting on the fix. So I started reaching out. Zero ad spend, all cold outreach, no platform, no funnel — just telling people directly that the tool existed and letting the results argue for themselves. Veridian is now used by 2 schools and 70 independent students preparing for the same exams I was.

None of that made the selling part easy. Cold outreach for a tool nobody has heard of, built by someone who'd only just sat the exam himself, is mostly rejection with the occasional yes buried in it — and the yes only starts coming once enough people have used it that the results can do the talking instead of me. That's still the actual mechanism: not a pitch, a track record other people are willing to check.

There was also a quieter shift the moment the first outside student used it: the tool stopped being a mirror and started being a judgment call I was making on someone else's work. When Veridian marked my own paper, I was the only one who paid for a mistake in its confidence flags. Once schools and strangers were relying on the same grading logic to tell them where they actually stood, the margin for the tool being wrong wasn't mine to absorb anymore — it was theirs. That's the real difference between building something for yourself and building something to sell: the standard doesn't get to slip just because you're no longer the one it would hurt first.

It's expanding to the SAT next, then more subjects, more boards, and eventually lower grade levels — the same build-it-for-yourself-first approach, applied to a wider market each time. The thesis hasn't changed since the first version: build the tool you wish existed, prove it on yourself before you ask anyone else to trust it, and only then let the market decide whether it was worth scaling. Most tools fail the first test. The ones that pass it tend not to need much convincing after that.