Tampa solo founder builds AI research platform that exposes when AI models disagree
MelMat asks eight AI models the same question and prints where they disagree. One recent brief came back with 51% consensus, which is rather the point.
MelMat, a Tampa research platform, sends the same question to eight AI engines and reports where their answers fail to line up.
Vanja Todorovic built it alone. She spent more than 25 years in financial services before leaving the corporate world, and says she operates MelMat without a traditional software engineering team.
The company says development began in March 2026 and the first public version launched on 1 May, roughly six weeks later.
The premise comes from compliance work. Todorovic's account is that a confident answer is not the same thing as a verified answer.
So MelMat does not pick one model. Multiple independent engines research a question in parallel, and a synthesis layer turns their work into a structured brief.
That brief is meant to expose four things: where the engines agree, where they contradict each other, the sources and grounding behind important claims, and the questions nobody resolved.
It also keeps the individual model perspectives rather than collapsing them into a single voice.
The company is explicit that consensus is not proof. It cites one recent brief that returned 51% engine consensus and partial grounding, published as such instead of being smoothed into one confident reply.
MelMat has since added M\u00b2W, described by the company as persistent research workspaces for longer investigations. Users accumulate research, facts and run history, then keep working against that context, compare subjects and produce summaries and memos.
The company is bootstrapped and says it is patent pending. It has been accepted into Embarc Collective, the Tampa startup hub.
The name is domestic in origin. MelMat comes from the names of Todorovic's twins, Melania and Mateo.
The bet underneath the product is that answers are getting cheap and verification is not. If any model can produce a plausible paragraph in seconds, the scarce thing becomes knowing which paragraph is supported and which one is a guess wearing a suit.
For founders, that is a practical question rather than a philosophical one. Anyone using AI output in a pitch deck, a diligence pack or a regulatory submission is already carrying the risk that the confident answer was wrong, and eight engines disagreeing in writing is at least a record of that risk.
Pricing has not been announced, and the company has not released user or revenue figures.
More: MelMat
