The only person I can actually quote on this story is an aerospace analyst named Scott Mikus, who appears in a Howmet Aerospace earnings call asking about F-35 fleet utilization. He is introduced as being with Melius Research. He is not, as far as anyone can tell, involved in a $20 million funding round for a startup founded by former Ramp engineers.
That is where my research trail ended, and I think that is the more useful story to tell you today.
What I was asked to verify
The premise: ex-Ramp engineers raised $20 million for a platform called Melius, after killing their first product and starting over. It is a clean narrative. Fintech pedigree, a public pivot, a mid-size seed or Series A. The kind of thing that gets 400 upvotes and a dozen reposts before anyone checks the cap table.
What my sources actually returned:
- A podcast feed description for a daily business and finance show that recycles Seeking Alpha headlines
- A German-language stock page asking whether to buy Howmet Aerospace
- An earnings call transcript where an analyst from Melius Research asks a question about defense spending and capital expenditure
- A Reddit bot-check page telling me to prove I am human
Not one of those documents mentions a funding round. Not one mentions a startup. The word “Melius” appears exactly once, attached to a sell-side research firm that has existed for years and has nothing to do with a pivot story.
The name collision problem is going to get worse
This is the part worth sitting with. “Melius” is Latin for “better,” which makes it catnip for naming consultants. There is already a Melius Research in equity analysis. If a new company adopts the same name, every automated retrieval tool pointed at the keyword will pull earnings call transcripts for the rest of time.
I review AI tools for a living, and the failure mode I see most often is not hallucinated prose. Modern models have gotten noticeably more careful about inventing paragraphs. The failure is upstream: the retrieval layer grabs documents that share a token with your query, the ranking layer has no idea that a research boutique and a seed-stage startup are different entities, and the summarizer gets handed four irrelevant files and a CAPTCHA.
A model that writes beautifully on top of garbage retrieval is worse than a model that writes clumsily on top of good retrieval. The second one you can edit. The first one you have to fact-check line by line, which costs more than writing it yourself.
The Reddit wall is a real signal
One of my four sources was literally a bot-detection page. That is not a quirk. Large chunks of the social web have closed to automated readers over the past few years, and startup funding chatter lives disproportionately in exactly those closed places: Reddit threads, X replies, private Slack groups, LinkedIn posts that never get indexed properly.
So when a research tool reports back with nothing, you cannot tell whether the event did not happen or whether the evidence sits behind a gate your scraper cannot pass. Those two outcomes look identical in a results pane. Any tool that does not distinguish between them is handing you a confidence level it has not earned.
What I would want from a tool that handled this correctly
A solid research agent should have told me, in plain language, that it found a name match on an unrelated entity and could not confirm the funding event. Instead the raw output was a pile of documents with the implicit suggestion that I make something of them.
The useful features here are unglamorous:
- Entity disambiguation, so a company name maps to an identifier rather than a string
- Explicit gap reporting, where the tool says what it could not reach and why
- Source-type labeling, so a podcast blurb does not sit at the same weight as a primary filing
- Blocked-request transparency, so you know a CAPTCHA ate your coverage
Where this leaves the story
The Melius funding round may well be real. Ramp has produced plenty of engineers who went on to start companies, and scrapping a first product before raising is an ordinary thing that happens constantly and rarely gets written up honestly. If someone has the filing or an on-record statement from a founder, I would genuinely like to read it.
But I am not going to write 800 words of texture around an event The tell is specificity without sourcing: a precise dollar figure, a vague “people familiar with the matter,” no filing, no founder quote, no investor name.
If your AI research stack produced a confident draft of this story, go check what it was actually reading. There is a decent chance it was an aerospace earnings call and a request to prove it was human.
🕒 Published: