Refery Engineering Documentation

Refery is an AI hiring intelligence engine for venture-backed startups, built on a hybrid retrieval, scoring, and adversarial evaluation architecture. This documentation describes the production systems that power candidate-role matching, pipeline orchestration, and the human-in-the-loop scout network that generates Refery's training signal.

These docs are written for engineers, partners, hiring managers, and evaluators who want to understand how Refery actually works under the hood. Nothing here is marketing copy. Every algorithm, schema decision, and engineering principle described is in production today.

Reading order

1. Architecture overview — system diagram, tech stack, data flow
2. The matching engine — multi-vector retrieval and the signal engine
3. The evaluation panel — five-persona adversarial scoring with hard veto
4. The pipeline engine — state machine, reconciliation, append-only history
5. Rule engines — declarative job filtering, tier waterfall outreach
6. The data flywheel — scout network as distributed HITL labelers
7. Scalability and efficiency — why Refery uses ~99% less compute than naive AI recruiters
8. Roadmap — what is shipping next

Engineering principles

Refery's core architectural bet is that most hiring decisions are deterministic and only a small set are genuinely ambiguous. The platform encodes that bet directly:

• Deterministic signals (logo tier, trajectory, comp reach, visa fit) are computed by rules, not LLMs.
• Embedding-based retrieval narrows millions of possible candidate-role pairs to a few dozen high-signal candidates per role.
• LLM compute is reserved for the genuinely ambiguous final step: the five-persona panel. This is where Refery spends its tokens.
• State transitions are evidence-bound and append-only. The pipeline engine never moves a row backward without explicit human override, and every transition leaves an auditable trail.
• The scout network of 300+ operator partners is treated as a structured labeling layer. Every placement outcome refines the scoring model.

The result is a system that produces top-1% recruiter quality output while consuming roughly 1% of the compute that a naive "LLM-evaluates-every-pair" architecture would burn through.

Status

Production. Refery operates an active pipeline of senior engineering and GTM placements at venture-backed startups (seed through Series B) primarily in the US (SF Bay Area, NYC) with the platform itself operated from Barcelona, Spain.