All work

AI first-round phone screening

Screenline

Recruiting firm, India

A recruiting firm in India wanted to take the repetitive first screening call off its recruiters' plates without handing hiring judgement to a machine. We built Screenline, a private recruiter workspace. A recruiter uploads a job description and a resume, reviews an AI-prepared brief and the candidate's phone number, and only then confirms a call. A voice agent phones the candidate using native realtime audio: it confirms it is speaking to the right person, says it is an AI, asks permission, and then works through the approved topics and practical questions such as notice period and availability. After the call, the recruiter sees each answer marked as self-reported, uncertain or unanswered, with the exact candidate quote behind it. Over 12 days the work moved through 19 scoped tickets to a pilot that is safe by default: live calling stays blocked until every approval, cost cap and provider check is in place.

screenline
Recruiting firm, India

At a glance

Client
A recruiting firm in India running an English-first hiring pilot, with international numbers added later
Problem
First-round phone screens take up recruiter time, and the notes they produce are inconsistent and hard to check against what the candidate said
Solution
A restricted workspace that runs reviewed, consented AI voice screening calls and returns quote-backed results for a person to decide on
Timeline
12 days, 19 tickets, from first commit to a working pilot
Engineering
62 commits, 9 database migrations, 42 test files (113 test cases) plus 12 runtime and provider check scripts
Stack
Next.js 16TypeScriptPostgreSQLBetter AuthOpenAI RealtimeExotelWebSocketslibphonenumber

The challenge

Phone screening is a good fit for a voice agent, but it is also where automation does the most damage when it gets something wrong. A call to the wrong person, a call nobody approved, or a summary that invents a skill all land on real candidates and on the firm's reputation.

No accidental calls

Uploading or parsing a resume could never start a paid call. Every call needed an explicit, reviewed and confirmed action.

Telephony is unforgiving

Real-time audio, interruptions, socket drops, retried webhooks and ambiguous timeouts all had to be handled without double-dialling anyone.

Identity and consent first

The agent confirms it has the right person before mentioning the resume, says it is an AI, and respects a refusal immediately.

Results must be checkable

No scores, no "verified skill" labels, and no summary prose the candidate never said.

Hard limits on spend and hours

Per-call duration and cost, total pilot spend and calling hours all enforced, with no defaults made up in code.

A safe place to rehearse

Recruiters wanted to try the agent on realistic test candidates without touching real candidate data.

Our approach

We split the system into a Next.js web app for recruiters and a single persistent worker that owns the job queue, the audio bridge and data retention. Each capability arrived as its own ticket on its own branch, with written acceptance evidence recording what was proven locally and what still needed live verification. The code enforces that distinction too: live calling is gated behind explicit approval flags, and a readiness panel shows every remaining blocker on screen.

What we built

1

A restricted recruiter workspace

  • Public signup disabled; every API checks company membership and returns 404 for another company's records.
  • Jobs, Candidates, Screening Results and Test Call in one workspace.
  • A readiness panel checks the database and worker heartbeat and lists each configuration and safety blocker in plain language.
2

Resume and job-description intake with human review

  • JDs as text, PDF or DOCX; one private resume per candidate application.
  • Parsing runs in an isolated process with a timeout and memory cap.
  • An AI brief lists requirements, resume claims (each tied to a resume and JD quote), gaps and topics. Any quote not found in the source is rejected.
  • The recruiter confirms identity, phone number and topics. Replacing a document invalidates the review.
3

Reviewed, one-at-a-time calling

  • A database lock and unique index allow only one active call at a time; idempotency keys stop double submissions.
  • Eligibility is checked at submission and again just before dialling: review freshness, opt-outs, JD expiry, calling hours and unchanged inputs.
  • An ambiguous telephony timeout triggers read-only reconciliation, never a blind redial.
4

A native-audio voice agent on the phone line

  • The worker bridges the telephony media stream straight to a realtime speech model, with no text-to-speech chain in between.
  • Natural interruptions: the agent stops, clears playback and keeps track of what the caller actually heard.
  • Identity check, then AI disclosure, then a willingness question, all before any screening. Documents are passed in as untrusted data.
  • The model gets only narrow tools, and ending a call for opt-out or wrong number must cite validated candidate evidence.
5

Evidence-based screening results

  • Each topic, resume claim and practical question is marked self-reported, uncertain or unanswered.
  • Every answer carries exact candidate quotes, checked against the transcript in code.
  • Results record the model, config version and input hash, and call state is tracked separately from screening outcome.
6

Test Call: rehearsal without real candidates

  • Recruiters generate a paired synthetic JD and resume, both editable.
  • Browser simulation uses the same realtime session and policy, with a single-use, 60-second token and no spend.
  • A comparison view lines up each requirement with resume quotes and candidate answers, never ranking the candidate.

Security and operations

  • Live calling is off by default and needs separate approval flags plus sponsor-set duration, cost, spend and calling-hour limits.
  • Provider keys stay on the server; media connections need authentication plus a short-lived, single-use signed token.
  • Telephony webhooks are checked by source and shared secret, deduplicated, and can only move state forward.
  • Resumes and JDs sit in private storage, and a worker deletes expired documents and conversation evidence on a set retention period.
  • Logs are allowlisted and sanitised: no documents, transcripts, audio, phone numbers or emails.
  • Opt-outs and wrong numbers write suppression records that are checked before every dial. Hiring decisions stay with the recruiter.

Results

An end-to-end pilot path

Intake, review, confirmed dialling, a natural voice conversation and quote-backed results run in one system, and live test calls have connected end to end.

Safe by construction

One active call, no blind redials, cost caps and approval gates are enforced in the database and code, not left to operator discipline.

Evidence a recruiter can check

Every screening note links to the candidate's own words, and gaps show as "unanswered" instead of weaknesses.

Honest readiness

The workspace separates what is proven from what still needs live verification, giving the client a clear checklist before rollout.

Key takeaways

  1. 1

    Put the gates in the data layer. Unique indexes, locks and idempotency keys did more for call safety than any confirmation dialog could.

  2. 2

    Make AI output extractive. Requiring every answer to cite an exact transcript quote kept summaries trustworthy and easy to audit.

  3. 3

    Build to the provider's real behaviour. Live test calls exposed behaviour the documentation did not, and each defect became a regression test before it was fixed.

Have a process like this?

Tell us what is costing your team the most time. We will show you how we would scope it, build it and get it into production.