IENA matching engine

The problem

IENA places international applicants at summer camps across the US. The organization handles more than 100,000 applicants across more than 1,500 camps.

Every year, staff matched applicants to camps by hand. That is slow, it depends on who is doing the matching, and mistakes affect real placements.

What I did

I built the recommendation engine that suggests the best-fit applicants for each camp, so staff start from a ranked shortlist instead of a blank page.

• Semantic profiles. Applicant profiles and camp descriptions are turned into vector embeddings with OpenAI, so matching works on meaning rather than exact keywords.
• Vector search. PostgreSQL with pgvector finds the camps and applicants that sit closest to each other.
• Explained matches. GPT-4 writes a short explanation of why each recommendation is a good fit, so staff can check the reasoning instead of trusting a score.
• Tests. The engine has full unit test coverage with Jest.

The result

The engine went live in production and streamlined IENA's annual matching cycle: staff started from ranked, explained recommendations instead of reviewing every profile by hand.