from osmoda research · recruiting
843 sourced, 47 screened, before the recruiter logs on.
frog runs the funnel overnight, audit-ready. LinkedIn + GitHub + Wellfound sourcing, ICP scoring with applied rubric, Twilio voice screen with on-device Whisper, ranked into Greenhouse/Lever/Ashby — and every score in a NYC-AEDT and EU-AI-Act-ready ledger.
TL;DR
- • frog sources LinkedIn + GitHub + Wellfound, scores against the ICP, runs Twilio or Telegram voice screens, ranks for the recruiter
- • US average cost-per-hire is $4,700 with a 44-day time-to-fill; tech roles ~$6,200 [SHRM 2024]
- • LinkedIn Recruiter Corporate runs $10,800–$15,000/seat/yr and still demands ~7.3 hrs/recruiter/wk in active sourcing
- • NYC AEDT, Illinois HB3773, EU AI Act all require disclosure + bias audits — frog logs every score to a SHA-256 ledger
- • Overnight run, $3.40 in LLM + voice cost. Hands off to Greenhouse / Lever / Ashby via API
1. The pain — sourcing as labor
SHRM benchmarks the average US cost-per-hire at $4,700 and time-to-fill at 44 days; tech and engineering roles run closer to $6,200. The bulk of that cost is recruiter time, not job-board spend. LinkedIn Recruiter Corporate sits at $10,800–$15,000 per seat per year, and recruiters using it report ~7.3 hours per week — about $13,900/year per recruiter — just inside the search-and-filter step before any human contact. Effective InMail response rates run 10–25%, and only 4.77% in SaaS, so the funnel is deep before screening starts.
The legacy stack is a license stack. LinkedIn Recruiter ($10k+/seat/yr) for sourcing. Greenhouse ($6–25k/yr), Lever ($4–20k/yr), Workable ($5–12k/yr), or Ashby ($4.8k–$70k/yr) for the ATS. Gem layered on top for sequencing. Phone screens are still recruiter-time, recorded by hand. Voice screening tools (HeyMilo, Glider, PhoneScreen AI, Talkpush) exist but live outside your audit and your ATS.
It's getting worse — and stricter. NYC Local Law 144 (AEDT) requires annual bias audits and applicant disclosure for any automated employment decision tool, with $500–$1,500/day fines. Illinois HB3773 governs AI video interviews. The EU AI Act (Reg. 2024/1689, in force August 1, 2024) classifies recruiting AI as high-risk; full obligations enforce on August 2, 2026. "We just used ChatGPT" is no longer a compliance posture.
LinkedIn Recruiter
Industry-default sourcing seat at ~$10k+/yr; still requires the recruiter to read profiles and write InMails by hand.
Greenhouse / Lever / Ashby
Best-in-class ATSes; great for tracking, not for sourcing or for running the screen itself.
Gem
CRM/sequencing layer over LinkedIn; automates sends but doesn't conduct or score the screening conversation.
HeyMilo / Glider / PhoneScreen AI
Standalone voice-AI screeners; live outside your ATS, no shared audit ledger, point-tool sprawl.
2. The workflow — sourced, scored, called, ranked
- 1 · soot pulls a fresh role spec from Greenhouse/Lever/Ashby, expands it to a structured ICP (skills, seniority, comp band, geo, work-auth) via tool call.
- 2 · frog queries LinkedIn (via your seat), GitHub (search API + commit graph), and Wellfound; scores against the ICP and dedupes against ATS history.
- 3 · naga applies AEDT-compliance gates: applicant-class blinding, fixed scoring rubric, jurisdiction check (NYC/IL/EU); records the applied rules to ledger.
- 4 · frog sends personalized first-touch via Telegram, WhatsApp, or email; on consent, places a Twilio voice screen with on-device Whisper transcription and a structured rubric scorer.
- 5 · lantern packages a ranked shortlist with transcript, score, rubric, and audit hash; pushes to Greenhouse/Lever/Ashby as candidates with notes; flags ambiguous calls for human review.
3. Why it works
Compliance is the product
Every score, every prompt, every model version is hash-chained in a SHA-256 ledger that satisfies the NYC AEDT bias-audit data requirement, the EU AI Act high-risk record-keeping requirement, and Illinois HB3773's consent-and-disclosure trail. naga blocks any signed action — outreach, score write, ATS update — that doesn't pass the active jurisdiction's policy.
Voice is feasible now
Twilio Voice handles dial-out, on-device Whisper transcribes locally (no third-party transcript leakage), a typed-tool scorer applies a fixed rubric, and the recording + transcript + rubric land in your ATS as an audit-grade artifact. Recruiters resist AI screening when it's a black box; this isn't.
Augment, don't replace
You keep your ATS and your LinkedIn seat. frog augments — it doesn't try to be Greenhouse or Ashby. The 92-tool runtime means a recruiting flow is a graph of typed calls (search, score, dial, transcribe, rank, write), not a giant unverifiable prompt.
FAQ
Are we allowed to do this in NYC?
Yes, under Local Law 144 if the tool is bias-audited annually and applicants are notified. frog produces the audit artifact and the disclosure copy; you publish it. The hash-chained ledger is the documentation evidence the bias audit needs.
Does it replace recruiters?
No. It replaces the search-filter-dial-summarize loop. Your recruiter walks in to a ranked list with transcripts and rubric scores, then runs the human round. The 4.77% InMail-response funnel becomes a screened shortlist.
Can a candidate opt out of AI screening?
Yes — Local Law 144 requires it. frog routes opt-outs to a recruiter queue with a flagged note, no penalty applied to the candidate's score. The opt-out itself is logged for the bias audit.
Run frog overnight, audit-ready, on your NixOS box or managed cloud — from $29/mo.
Spawn a recruiting agent →