A hard part of MessyData is defining when not to trust the AI interpretation. It works best with repeated patterns, not long prose or unrelated notes, so review, warnings and safe-use limits are part of the product. It quickly becomes a fuzzy boundary though!

Here is a concrete example of what MessyData does. This fictional input contains copied expense notes with inconsistent dates, separators and a missing amount. MessyData turns the pasted text into a table, highlights points to check, and exports it as CSV. messydata.app

A prototype doesn’t always need to prove an idea will work. Sometimes it just needs to answer the next question. - Will people understand it? - Can it actually be built? - Is the problem worth solving? Build enough to learn something, then decide what comes next on the path to success.

Building small software experiments means not every idea needs a huge commitment. Just build enough to make it real. Put it in front of people. And learn what happens next. Some ideas grow, some change direction, and some stop. All three outcomes teach you something.

One thing building MessyData has taught me: The first version of a product isn’t just there to validate the idea. It’s there to teach you how to explain it. The questions people ask often reveal more than the features they request.

The best feedback I’ve had building MessyData has been the questions I’ve received: “Does it handle PDFs?” “What’s the tech stack?” “Why not agents?” Good questions expose assumptions you didn’t realise you made. Addressing confusion can be more valuable than adding features.

One thing I enjoy about building small AI products is discovering where not to add complexity. The first version of MessyData is deliberately simple. Every feature has to earn its place. I think that’s easier to improve than starting with something clever but hard to understand.

Building MessyData has made me appreciate something: The product isn’t the AI. The product is everything around the AI. The prompts, the review step, the expectations you set, the places you deliberately don’t automate. That’s where trust comes from.

1/ I got a useful early question about MessyData already making me glad to have started this “build in public” effort: “Does it handle multi-column PDFs?” The answer: not reliably enough to claim (/yet). MessyData is currently a pasted-text cleanup tool, not a dedicated PDF parser.

1/ A closer look at the first live tool from AI Innovation Design Lab: MessyData. The problem is simple: Useful data often starts messy. Copied from websites, PDFs, emails, notes, spreadsheets, reports, or internal systems.

Starting to build more openly around AI Innovation Design Lab. Focus: practical AI products, prototypes, and workflow tools. First live tool: MessyData, for turning messy pasted text into cleaner structured data. I’ll share what gets built, what works, what breaks, and where AI actually helps.