Gallery
What it does
Companies and researchers rarely find each other, and when an AI tool suggests matches, it often makes up reasons that sound convincing. Shark Tank for Science fixes that by putting every match through a pitch room. Four scout agents search the EINSTEIN dataset in parallel and suggest researcher and company pairs. Two shark agents, a skeptical CTO and a strict professor, question each pair at the same time, and a critic agent removes any claim that doesn't cite a real paper or company record. The pairs that survive are turned into ready-to-use joint project proposals that list the risks and rebuttals.
How the agents run concurrently
All agents join one shared Mozaik runtime and coordinate only through typed events on a shared state board. That board holds the candidate pairs, the claims and the verdicts. Four scout agents (company problems, new papers, researchers with no industry ties, cross-field researchers) search the dataset at the same time and post pairs to the board as they find them. When a pair lands, the CTO shark and the professor shark question it in parallel while the scouts keep searching. The critic watches every new claim and removes any it can't trace to a dataset record. The writer starts drafting a proposal as soon as a pair survives, so writing overlaps with reviews still going on. A chair agent applies the survival rule and a recorder writes the JSONL trace, and neither calls a model. The trace's timestamps show the scouts, sharks, critic and writer running at the same time.