AI Headhunter

End users are recruiters and agencies running talent search. Headhunter takes a thin intent and progressively enriches it into structured candidate rows — discovery across multiple sources, a relevance gate on the current title, verified email, and a deduplicated CRM-linked row per person. Production works, but it leaks money and forgets what it finds: every run goes straight to paid sources even when half the answer is already in the database, the sources are hardcoded so adding one is surgery, and results land in two tables that never reconcile. 2.0 fixes all three with one idea: the internal database is not a destination, it is the first rung of a waterfall - and every run feeds it.

Two designs, one toggle. Production today is the shipped six-step pipeline: a host-side orchestrator drives Strategist → Storage, reaching providers only through key-holding source tools, with supplements that fail open. The 2.0 design re-arranges the same job around the people intelligence graph: internal-first recall, an agent that routes only the gap, and a write-back flywheel. Toggle the two, then click any box or node to inspect it.

Scope is one LLM pass: it turns the brief and UI filters into a typed SearchSpec - the what, never the how. It sets the one decision that branches the whole run: mode. Pick a brief and watch it resolve; hover any spec field to trace it back to the tokens it came from.

Companies don't have to be typed. A company_scoped run pulls them from four places, and provenance rides along. resolved sources (CRM, Sonar) arrive with a domain and an entity id, so Scope passes them straight through; raw sources (manual, CSV) arrive as a name Scope has to resolve. Click a source to see the companies[] entry it yields.

• resolved · CRM
• resolved · Sonar (+ signal)
• raw · manual / CSV

This is what "internal database first" actually buys. A brief wants 100 people. The graph answers some for free; only the gap waterfalls out to paid sources, cheapest first, stopping at the first hit. Drag how many matching people already live in the graph - or load a scenario - and watch the bill move.

Production hardcoded each source and hand-wired it into the pipeline - adding one was surgery. 2.0 puts every source behind one contract, and the agent reads that metadata to route. The order below is the waterfall priority the planner follows. Click any source to expand it.

The qualifier is the one place precision is won or lost — production shipped it, and 2.0's Qualify step keeps it unchanged. It scores each candidate's current title against the search intent — a free substring pass first, an LLM synonym pass for the rest — and drops anyone below 40. Pick a search and a candidate and watch the verdict. Prior roles never count.

Bands and the synonym map here are illustrative of the deterministic-then-LLM code path. The point: the gate judges the current role, the prior role is preserved for context but never lifts the score.

Every candidate - internal or freshly found - runs the same identity cascade before it touches the graph. Deterministic keys first, then similarity, then semantics. The first tier to match wins; matches merge by field-union, no human review. Pick a record pair and watch which tier resolves it.

Production splits discoveries across workflow_run_people and crm_people - two tables that never reconcile. 2.0 unifies them into one graph that read and write-back both target. Four layers, one canonical person.

The forks settled in the design session. Each one trades a knob for less friction - the 2.0 bet is that automatic beats configurable for this workflow.

Graph first - nothing works without the unified store. Then resolution, then the internal rung for an immediate cost win, then the registry, then write-back to close the loop.