Finding U.S. partners beyond a Japanese startup’s industry

A medical-device startup describes “microneedle delivery”; an agricultural buyer asks for “precision nutrient injection.” Keyword searches can miss the shared action: penetrating a material to deliver a controlled dose. The proposed tool borrows from historical linguists, who use regular sound changes to trace different words to a common ancestor. It learns two sets of translation rules: one from past matches to account for understatement and promotional language, and another from patents that describe the same mechanism in different industries. It reduces each capability to four parts: the physical action, the material it acts on, the result, and measurable operating limits such as depth and force. It then generates search phrases in potential buyers’ terminology while preserving those limits. Each phrase is tested against real procurement calls and other buyer-side documents; phrases that find no evidence are dropped. Scouts receive ranked opportunities with the shared mechanism and supporting documents, so they can assess physical fit rather than rely on keyword overlap or an unexplained similarity score.

A firm can announce a new direction without knowing which capabilities it needs to get there. We map its current capabilities and declared goal into a network of functional steps, then look for missing links, just as biologists use GapFill to find missing reactions in a cell’s chemical network. We describe Japanese startups by what their technology does, rather than their sector, and consider only gaps backed by independent demand evidence, such as a procurement request. For example, a micro-needle startup becomes a candidate for soil-nutrient delivery because both require controlled penetration and delivery of a measured dose at depth. Across the network, we first maximize the number of firms whose paths can be completed, then choose the fewest startup connections that achieve that coverage. This also reveals startup capabilities that fill gaps for several firms. Removing each proposed connection and recalculating shows whether it is necessary within the modeled network, replaceable only with more connections, or interchangeable. Scouts receive a short list of introductions with the missing step and supporting evidence, rather than a ranking based on similar descriptions.

Scouts can spend hours investigating one partnership, yet face millions of possible pairings between Japanese startups and U.S. partners. Keyword search misses opportunities when a technology and a need share a working principle but use different words. We borrow the way chemists choose their next costly experiment. First, multilingual AI extracts what each startup’s technology does, how it works and under what constraints, alongside comparable descriptions of partner needs. A predictive model estimates each pairing’s fit and the uncertainty of that estimate. We rank pairings by predicted fit plus an uncertainty bonus, so unfamiliar cross-industry possibilities get a place alongside familiar, promising matches. Scouts receive a short explanation of the shared mechanism, then investigate commercial fit. They record both successes and failures to update the model. Instead of repeatedly searching familiar vocabulary, scouts build evidence about which mechanisms transfer between industries and use it to choose the next investigation.