AI recruitment middleware that connects Salesforce to vector search
Recruiters drown in résumés
Keyword search misses strong fits. Manual review does not scale. We built AI-powered recruitment automation—middleware that sits between a Salesforce ATS and modern AI retrieval.
Pipeline overview
Salesforce webhook → FastAPI gateway → Embedding service → Pinecone query → Ranked matches → Proposal draft
MatchingService, OpenAIService, VectorDBService, and DocumentProcessor compose the core—each tunable without redeploying the whole stack.
Technical choices
| Component | Role | |-----------|------| | Azure OpenAI (gpt-4.1, text-embedding-3-small) | Embeddings and proposal text | | Pinecone | Vector store for résumés and job descriptions | | SQLite + SQLAlchemy | Operational state and assistant instructions | | FastAPI + Uvicorn | Fully async, high-concurrency API | | Docker | Cloud-ready deployment |
Matching blends semantic similarity with business rules: skills, experience, geography, compensation, and availability.
Why middleware beats monolith
Recruiters keep Salesforce. Engineering keeps AI velocity. A thin, composable layer means you can swap embedding models, tune prompts, and adjust ranking without migrating the ATS.
If you are building similar systems
- Start with webhook → embed → retrieve → rank as the smallest valuable loop
- Persist prompt and retrieval config in versioned storage—not hardcoded strings
- Ship a lightweight ops dashboard early; non-engineers will need to tune behavior
Need help designing AI middleware? Talk to us.