back

Appairium

A natural-language search engine matching professionals with over 25,000 tools via semantic vector search.

Appairium search interface showing natural language query results.

Overview

A B2B matching platform that allows professionals to find the right software by describing their needs in natural language. The system ranks recommendations from a catalog of over 25,000 tools based on semantic relevance and constraints.

Why it matters

Traditional directory search fails when users cannot name specific categories or features. Appairium accelerates decision-making by understanding the intent behind a “messy” user prompt and mapping it to specific software capabilities instantly.

Key capabilities

  • Semantic Discovery: Users input raw problem descriptions (e.g., “I need to manage construction site payroll”) rather than keywords.
  • Smart Ranking: Returns the top 3 most compatible tools with a calculated compatibility score.
  • Constraint Filtering: Refines vector results with hard filters like price, currency, and integrations.

Technical approach

  • Vector Search Pipeline: Utilizes Pinecone to manage embeddings for the massive product catalog, ensuring scalability and speed.
  • Hybrid Reranking: Implemented a two-step ranking process: initial vector retrieval followed by a cross-encoder reranker to handle ambiguous prompts.
  • Performance: Optimized for sub-3-second latency from prompt to ranked results.

Role

Co-founder and AI Lead. Built the data preparation pipeline, vector database infrastructure, and the ranking/reranking algorithms.

Outcomes

  • Performance: Achieved high relevance on complex queries with near-instant results.
  • Exit: The project was successfully acquired.

Tech stack

Python, Pinecone, Vector Embeddings, Cross-Encoders