AgenticHire

Inspiration

Hiring is supposed to find the best person for the job. In practice, it finds the best-connected person who formatted their CV correctly.

88% of employers admit their ATS filters out quality candidates. Referrals dominate because they're cheap, fast, and carry assumed quality — but they're also socially exclusive. The candidate pool is shaped before a single interview happens.

When I built my card reselling business and working at startups Bumpp and Ecovolt, I saw the problems that hiring brings and decided to make a solution to solve them for employers as well.

What it does

AgenticHire gives every applicant — regardless of school brand, referral, or network — the same structured AI interview.

Employer posts a role — paste a JD, answer a few questions about what "good" looks like at 90 days. Under 5 minutes.
AI interviews every candidate — a 10–15 minute async text interview, calibrated to the role. No scheduling. No interviewer bias. Consistent bar for everyone.
Employer reads the evidence brief — a hire/no-hire verdict with a 9-dimension capability score (Technical Depth, Problem Solving, Ownership, Communication, Resilience, and more), backed by verbatim quotes from the interview.

Capability decides. Not connections.

How we built it

AgenticHire runs on a multi-step agent pipeline — not a single LLM call with a long prompt.

Pipeline: JD Input → Question Bank Generator → Interview Agent → Scoring Engine → Evidence Brief Synthesiser

Question Bank Generator parses the JD and calibrates questions against benchmarks for similar-stage companies and roles. Questions are role-specific, not generic.
Interview Agent conducts the async conversation, probing follow-ups based on candidate responses.
Scoring Engine evaluates responses across 9 dimensions using structured rubrics — not raw LLM output. Scoring standards adjust from historical hire/no-hire feedback over time.
Evidence Brief Synthesise produces the final report with verbatim quotes and a calibrated verdict.

The system is currently calibrated across 10+ specific job types including software engineer, AI/ML engineer, and operations roles.

Challenges we ran into

Calibration consistency — Getting the scoring rubric to produce stable, comparable scores across candidates for the same role took significant iteration. A score of 7/10 on "Ownership" needed to mean the same thing every time.

Candidate friction vs. signal tradeoff — A longer interview extracts more signal but risks drop-off. We landed on 10–15 minutes as the ceiling after feedback from multiple founders who flagged that anything longer creates friction for applicants.

Preventing gaming — Structured text interviews are gameable if questions are predictable. We invested in question variance logic so candidates cannot pattern-match to a fixed set of answers.

Accomplishments that we're proud of

End-to-end working pipeline: JD in, evidence brief out, same day.
Validated the problem through discovery conversations with founders, HR leads, and operators across Singapore.
Built a scoring engine that produces 9-dimension breakdowns with verbatim quote attribution — not just a number.
Positioned against enterprise ATS incumbents with a zero-friction, no-integration-required setup that takes under 5 minutes to deploy for a new role.

What we learned

The hiring problem is not primarily a filtering problem — it's a surfacing* problem. ATS filters are not broken because they filter too much. They're broken because they filter on the wrong signal (keywords, school names, referrals) rather than demonstrated capability.

We also learned that the biggest objection from employers is not cost — it's trust(another reason why so many companies use referrals). They want evidence, not scores. That's why we built the evidence brief around verbatim quotes first and scores second.