You type the name into a search bar. Ex-Ramp engineers, $20M, a platform called Melius, a scrapped first product. It has the shape of a real story — the fintech pedigree, the pivot, the round size that’s big enough to matter and small enough to be believable. You hit enter expecting a funding announcement, a founder blog post, maybe a TechCrunch write-up with the usual quotes about conviction and iteration speed.
Instead you get an earnings call transcript. “Next question comes from Scott Mikus with Melius Research.” He’s asking Howmet Aerospace about F-35 fleet utilization. Below that, a podcast episode summary telling you to subscribe to Seeking Alpha Premium. Below that, a Reddit page that isn’t a Reddit page at all, just a wall asking you to prove your humanity.
That’s the entire evidence pile. No round. No founders. No product, scrapped or otherwise.
Why this matters more than the story itself
I review AI tools for a living, which means I spend a lot of time watching systems generate confident output from nothing. This is a textbook case, and it’s worth sitting with before anyone writes the version of this article that pretends otherwise.
The failure here isn’t exotic. It’s a name collision plus a paywall plus a bot gate. Melius Research is an established equity research shop — their analysts show up in industrial earnings calls, asking real questions about defense spending and capital expenditure. Any retrieval system searching for “Melius” is going to find that, because that’s where the indexed, crawlable, publicly available text lives. A startup that may have announced a seed round has a fraction of that footprint, and whatever footprint it does have is probably sitting behind a login or in a post that got gated.
So the system returns what it can reach. And what it can reach is an aerospace question about F-35s.
The part that should worry you
None of those three sources contain a single fact about the topic. Not a date, not a dollar figure, not a name. And yet you could absolutely write 800 words off them. You’d mention the fintech-to-infrastructure pipeline. You’d note how Ramp alumni have become a recognizable founder archetype. You’d write a paragraph about how scrapping your first product is a sign of discipline rather than failure. You’d close with something about how the pivot is the new launch.
Every sentence would be plausible. Every sentence would be unsourced. And it would read better than what you’re reading now, because invented specifics are more satisfying than an honest shrug.
This is the actual thing that goes wrong with AI-assisted journalism, and it’s not hallucination in the dramatic sense. It’s the quieter version: a system that retrieves weak sources, recognizes they’re weak, and then generates around the gap using pattern knowledge instead of saying the gap exists. The output is fluent. The citations are real documents. They just have nothing to do with the claim.
What a good tool does here
If you’re evaluating research agents or AI news tools, this exact scenario is your test case. Hand it a topic where the sources don’t support the premise and watch what happens.
- A solid tool tells you the retrieval failed and names what it actually found
- A weak tool writes the article anyway and footnotes the Howmet transcript
- A genuinely useful tool distinguishes between “no information exists” and “I couldn’t reach the information” — those are very different problems with different fixes
- The worst ones treat a bot-verification page as a source rather than as evidence that they got blocked
That last one is the tell. A Reddit human-verification challenge appearing in a source list means the crawl hit a wall. It is not a data point. Any system that passes it through as context has no concept of what a failed fetch looks like.
Where that leaves the story
Maybe ex-Ramp engineers did raise $20M for something called Melius after killing their first product. It’s a reasonable thing to have happened. Fintech engineers leave for infrastructure startups constantly, seed rounds that size get announced every week, and first products get abandoned more often than anyone admits publicly.
But I can’t confirm any of it from what’s in front of me, and I’m not going to write a profile of a company I can’t verify exists in the form described. If you came here for details about the round, the right answer is that this piece can’t give them to you, and neither can the sources it was built from.
The useful takeaway is the smaller one. When an AI research tool hands you three links and a confident summary, open the links. Check whether the Melius in the citation is the Melius in the headline. That five-second check is currently the difference between reporting and generating.
🕒 Published: