A researcher went looking for a problem
Most startups begin with a problem and go looking for the skills to solve it. Angie Simmons did it the other way round.
She is a researcher by training — computer science and AI research, with a PhD in IT — and she started not with a product but with a question about where her particular expertise would actually be worth something.
“I am a researcher by background, so it was more around finding the problems that could benefit from the area that I was working in.”
So she went and asked. Not a survey, not a market report: a long run of conversations across unrelated industries, looking for the overlap between what her team knew and what someone was actually struggling with.
And so this involved interviews with people in lots of different verticals until we found some where what the expertise I and my team had aligned with the real-world challenges that they were facing.
Where the overlap turned out to be
It was taxation law.
Hashtag.AI's first major customer is building AI for tax — a field where an answer is either grounded in the right document, from the right year, read in the right context, or it is worse than no answer at all.
“Our first major customer is doing AI for taxation law, so this is one of the places where you can't just take an answer out of context.”
That constraint is the product. A clause in a tax act does not mean anything on its own — it means something in relation to the rest of its document, to the other documents around it, and to when it was written.
You've got to take into account how it relates to the rest of the document, to the other documents. Is it even relevant, or is it something back from the 1970s?
Why not just use a bigger model?
The obvious answer to a hard question is a more capable model. Simmons' whole argument is that this is the wrong lever.
Structure the corpus properly and a modest model has what it needs. Leave it unstructured and the largest model available is answering from whatever it happened to pull back. The company's name for what it sells is deliberately plain: reliable knowledge for AI agents.
It's not so much about having the biggest, latest large language model like ChatGPT, Gemini or Claude. It's about making sure that the data is structured in the right way, that you can reliably retrieve the right information as context, to get a good answer.
The problem that looks easy
Retrieval has a reputation as a solved thing — a component you wire in rather than a field you research. Simmons is blunt about that.
It is also why her plans are not the usual ones. Instead of hiring to ship faster, she is looking for research partnerships and other researchers, on the view that several of the hard parts are genuinely unsolved rather than merely unbuilt.
Reliable AI retrieval sounds like an easy problem. It's not. There's so many open problems.
From a theoretical problem to actual customers
For a researcher, the risk is never running out of interesting questions. It is never leaving them.
What the AI Pathfinder program gave Simmons was not a technique. It was the push to keep having conversations outside the lab, and the effect of that was to move the work from something being reasoned about to something someone was paying for.
“So instead of being a theoretical research problem, we now have actual first customers and we feel a lot more motivated and directed about the research that we're doing.”
The word she keeps coming back to is connection — and specifically, knowing who to ask.
For myself though I found the most valuable part is the feeling of a connection. So if there's things that I'm uncertain about, I know how to contact.
Who this is for now
Hashtag.AI is looking for the same shape of problem it found in tax: an organisation in a regulated field, sitting on a large body of complex material, that has already tried the off-the-shelf tools.
That last clause is the qualifying one. The customers Simmons wants are the ones who have already discovered the problem is harder than it looks.
We're particularly interested to talk with people in regulated domains who might be exploring generative AI tools, particularly anyone dealing with large corpuses of complex data where they've tried the existing tools and they're not quite working.
