I Didn't Trust My Gut, So I Built a Tool Instead

I don't think “trust your gut” is good advice for decisions that are structurally unlike anything your gut has calibration for. That's the sentence that got LifePath started, somewhere around 1am, running the numbers on a decision I hadn't made yet and realising I had no actual data to run them on.

The spiral was an ordinary one: what if I'd taken a different path, different degree, different city, different first job. Everyone has that spiral. What surprised me wasn't having the thought, it was what happened when I tried to actually answer it properly instead of just feeling my way to a conclusion. I went looking for something that would tell me, with real data, what a career switch into consulting typically does to someone's trajectory over five years, or what relocating to another city actually costs against what people assume it costs. I didn't find it. I found Reddit threads full of survivorship bias, and I found AI chatbots that would confidently generate an answer that sounded specific but was actually just a plausible-sounding average of internet opinion, dressed up as analysis.

The difference between an answer and a sourced answer

That's the distinction that actually matters, and that most “AI-powered advice” tools blur on purpose, because blurring it is what makes the product feel magical. An AI model can generate an answer to almost any question you ask it, fluently and confidently, whether or not it has any real grounding in verified data. Ask it what relocating to Berlin does to your effective take-home pay and it will give you a number. Ask it twice and you might get two different numbers, both delivered with the same confidence. The fluency is real. The grounding, most of the time, isn't.

I didn't want a tool that generates a plausible-sounding opinion. I wanted a framework that forces the comparison to be built on something real: cost-of-living data that's actually sourced, salary trajectories that come from real reported outcomes, not a language model's best guess at what those numbers probably look like based on its training data. That's a much harder product to build, because sourcing and structuring “real data” for a question as genuinely messy as “should I move to Berlin” turns out to be a much harder problem than picking a model and writing a good prompt.

An AI model will always give you an answer. It will only sometimes give you a true one, and it usually won't tell you which situation you're in.

I tested this directly, early on, by asking three different general-purpose chatbots the same question: what does relocating from London to Berlin actually do to disposable income for someone on a similar salary. I got three different numbers, each stated with identical confidence, none of them citing a source I could go check. When I pushed one of them to explain its reasoning, it produced a plausible-sounding breakdown of rent, tax, and cost-of-living difference that, on closer inspection, was built from generic averages that had nothing to do with the actual comparison I'd asked for. That's not a hallucination in the dramatic sense people usually mean. It's something more common and more dangerous: an answer that's wrong in an unremarkable, hard-to-notice way, delivered with exactly the same fluency as a correct one.

The honest version of building LifePath so far has mostly been sourcing work, not model work. The actual product is deciding what counts as reliable evidence for a life decision, deciding how to structure comparisons so they're actually apples-to-apples instead of superficially similar, and deciding what to leave out because the data doesn't support a claim even though a chatbot would happily manufacture one. The interactive prototype and the Next.js and Supabase scaffolding underneath it are real, but they're not the hard part. The hard part is resisting the shortcut of just wrapping a language model around vague inputs and calling the output “personalised insight,” because that's exactly the failure mode I built this to get away from.

I think a lot of the current wave of AI products are, functionally, expensive ways of asking your gut a question and getting an answer that sounds more authoritative than a gut feeling actually deserves. That's not a knock on AI as a technology. It's a knock on treating fluency as if it were evidence. My actual position is narrower and more useful than “AI is untrustworthy”: it's that fluent-sounding output and sourced, verifiable output are different things, most products don't tell you which one you're getting, and the decisions worth building a tool for are exactly the ones where that difference is the whole ballgame.

I still don't fully trust my gut on decisions like this. I trust it slightly more than I trust an unsourced chatbot answer, and slightly less than I trust a framework that shows its working. LifePath is the attempt to build the third option, because as far as I could tell, in 2026, nobody else had.