What founder + Claude did by hand for 236 candidates — extraction, filters, ranking, probes, sequencing — now runs automatically on every submission. Humans keep the calls and the judgment.
Pipeline Stages
4 of 5 Automated
Intake to Score Time
< 60 Seconds
Replay Test Dataset
Real Candidate Data
Data Security
Firestore Rules + TLS
DECIDE stays human by design — the instrument informs judgment.
External redirect set at post creation disables platform questions
Automated1 question per screen, unmasked phone, rendered from hr_roles
AutomatedVerbatim extraction, hard filters, §3.4 formula, retention flags
AutomatedHumans keep the interviews and offers informed by generated probes
HUMANLoad-bearing Communications Guard blocks rejections until acceptance
AutomatedThe §3.4 ranking formula calculates candidates with exact arithmetic weights from their verbatim responses:
Single strongest discriminator: OWNED=1.0, COORDINATED=0.5, THEORY=0.0
Verbatim failure analysis: real failure=1.0 vs dodged=0.0
Specific metrics & consistency across funnel stages
Write-the-actual-message task. Has near-veto power on remote roles.
"Ask for exact metrics on 'I personally built candidate assessment workflow...'"
"Deep dive into 'Wrong assumption on notice period for senior candidate...'"
"Confirm monthly compensation expectation of ₹40,000/mo against current notice period."
Top 3 unchanged across 2 passes → Role frozen → WhatsApp: "list frozen — start calling."
The written answer is evidence; voice is verification. The AI call NEVER decides — delivery is strictly ADVISORY.
Evaluates Composure, Fluency, Energy, Warmth, and Structure on a 1–5 scale. Stored separately in a dedicated column.
Prompt hard-codes accent rules: Indian-English vocal features are NEVER read as emotional signals; delivery only, never content.
Alerts founder via WhatsApp ONLY when delivery contradicts written record (e.g. written top-3 candidate with delivery mean ≤ 2.5).
Connect to Firebase, run live assessments, and make confident hiring decisions.