Partner Intelligence Engine

Find and reach relevant importer-distributors of branded packaged food across a wide set of international markets — companies with a real, evidenced import capability for the product category, not just companies that mentioned relevant keywords online.

• The target market was fragmented and had low digital presence across most of the 32 countries.
• Total addressable market was relatively small, but each correct account carried high commercial value.
• Company information was inconsistent — categorization, language and registry formats differed by country.
• Conventional B2B databases returned a lot of noise: restaurants, chefs, food-service businesses, bulk commodity importers and suppliers that were not the actual target.
• Resources were limited — this had to run without a large team or an enterprise data budget.
• The output needed to be usable prospects with verified import capability, not raw, unqualified leads.

Apollo and Clay's standard data layers did not provide sufficient coverage for this ICP — the fragmentation of the market meant no single vendor owned complete data on it. Rather than force the problem into a standard database-first workflow and accept a high false-positive rate, I designed an alternative data and research architecture built around the market itself.

That meant treating company discovery and qualification as a research problem, not a lookup problem — combining directories, registries, trade and customs data, and reverse-engineering competitor distributor networks (a company already distributing a comparable product category is a far stronger signal than a keyword match).

The motion ran from signal to personalized outreach, with an explicit qualification gate before any contact was researched, and a human checkpoint before anything was sent. The system's job was to prepare research and personalization — not to replace judgment on relationship-sensitive actions.

Discovery and research required probabilistic reasoning — interpreting sources, classifying companies, verifying claims — that a deterministic workflow alone couldn't handle well. That's what moved the system toward an agentic architecture: eight agents with defined boundaries, inputs, outputs and handoff logic, coordinating through a shared data layer built on Supabase.

Qualification was intentional attrition, not failure — each stage filtered for evidenced import capability rather than keyword relevance.

On the contact layer, 159 decision-makers were researched, and 66 verified emails were found through a Clay → Apollo enrichment waterfall. Where a contact-level reveal failed, the system fell back to company-level targeting rather than guessing at a contact.

Separately, reverse-engineering known competitors' distributor networks surfaced 81 additional companies — each one already demonstrating category import activity, which made it a stronger signal than a keyword match from a database search.

• No single data vendor owns complete coverage for a fragmented, low-digital-presence ICP — the system has to build its own coverage.
• Competitor distributor-network research is a stronger discovery signal than keyword search for categories with established comparable players.
• Import capability needs to be verified through evidence, not inferred from marketing language.
• Agentic research earns its complexity when the task is reasoning-heavy (interpretation, classification, verification) — not as a default for every workflow.
• Keeping a human checkpoint before every send preserved relationship quality without slowing down research and qualification.