After 300+ manual thesis stats projects, I built the tool that replaces me
I run a statistics service for students. The same disaster kept arriving in my inbox: students paste data into ChatGPT, get plausible-looking p-values, and discover at their defense that the numbers were fabricated.
So I built Academic Stats Agent: upload your data, it picks the right statistical test, runs actual computations (scipy/statsmodels — zero AI in the stats engine), and generates an APA-7 report. The LLM only narrates numbers the engine computed; a fail-closed guard blocks anything fabricated.
Business model: analyses free and unlimited, paid credit packs (from €4.90) for formatted reports. No subscription. It sits between "free but you need to know stats" (jamovi, R) and "€99+ human statistician" — that middle was empty.
Try without signup: https://studentstats.net/tools
Full app: https://studentstats.net
Would love feedback on the pricing model especially.
The pricing thing isn’t really “how much should a credit cost?” it’s more “what panic moment is the student paying to get out of.” I’d try framing the paid step around the actual deliverable instead of the mechanism, something like “APA report you can defend” feels way stronger than “formatted report credits.” Keep the free analysis as a trust‑builder, then charge right when they actually need something to submit, share with a supervisor, or defend.
The distinction that stood out to me is that you're separating computation from explanation.
Most AI tools ask users to trust the entire output. Your approach asks them to trust the statistical engine first, then uses AI only to communicate verified results. That changes the trust model in a meaningful way, especially in an academic setting.
Thanks — "trust the engine first" is better than anything on my landing page, honestly. The surprising part was how little the AI can be trusted even for communication. Early versions would "improve" numbers while explaining them — round p = .047 to "about .05", call d = 0.3 "strong". Now the guard diffs every claim in the prose against the computed results and rejects the output on any mismatch. It fires often, which I've decided is the feature working. Curious if you'd see the same split working in finance or legal.
That's an interesting question.
I think the answer depends on one assumption that changes quite a bit between domains, so I'd rather not give you a one-size-fits-all answer.
I'd be happy to explain how I'd think about it in the context of your product.
What's the best email to reach you on?