private beta

Knock Recruit

From thousands of applicants to the ten worth your time.

Visit site

Applying for jobs is frictionless now. You can easy-apply, submit with AI, and fire off a hundred applications before lunch. One posting can pull two thousand of them. The recruiter drowns in text and can’t find the signal.

So Knock rewires the process and makes recruiting AI-native. Six agents read every applicant and rank them against every open role, so the recruiter only spends judgment on the ten worth judging. The match score is auditable math they can defend to a client — not a black box.

At a glance

Problem

A single job posting can draw 2,000 applications, and the incumbent tools just store candidates. They do nothing to help a recruiter decide who is actually worth attention.

Solution

I built a multi-tenant platform where six durable AI agents parse resumes, score video answers, and rank every candidate against every open role — surfacing a defensible top-ten shortlist with a transparent score breakdown.

Why it was hard

A recruiter has to defend a ranking to a paying client, so the score couldn’t be a black box. I let the LLMs produce signals under validated schemas, but kept the final composite as deterministic math — with confidence handling so a candidate who arrives with partial data is still scored fairly.

Impact

It’s deployed in private beta, turning raw applicant volume into a ranked, auditable shortlist. I hold a flat seat-based price by engineering the AI cost down instead of metering customers for it.

AI components

  • LLMs

    Claude Haiku for high-volume extraction and scoring, Claude Sonnet for reasoning (intake, culture profiling, conversation), with Gemini as an automatic rate-limit fallback.

  • Agents

    Six durable Inngest agents (intake, enrichment, scoring, outreach, briefing, compliance) plus a conversational Slack assistant.

  • RAG

    None — query-driven context, serializing top-N workspace entities into prompts for a bounded domain.

  • Evaluations

    Deterministic, auditable composite scoring with confidence gating and graceful degradation; an offline eval harness identified as the next step.

  • Workflows

    Durable, event-driven orchestration with SHA-256 input-hash dedup, concurrency caps, and months-long sleep-until quality-of-hire surveys.

  • Automation

    Job-board distribution (Indeed, Broadbean, Google for Jobs), cron-scheduled daily briefings, and an encrypted ATS sync.

Technology

Next.js • Supabase • Inngest • Claude • Gemini