Multi-Agent AI  ·  5 Specialized Agents

it calls. it interviews. it decides.

An autonomous AI that cold-calls candidates, conducts full telephonic interviews, and delivers a scored verdict - no human required.

Live Evaluation Output

Every call ends with a scored verdict - rubric-based, transparent, and fully automated.

HireLens evaluation result: Interview Complete with Accept verdict, score 128 out of 135, detailed category breakdown across correctness, communication, grammar, professionalism, and job fit, with strengths and weaknesses analysis

Actual output from a live HireLens session - Accept verdict, 128 / 135

The Agent Pipeline

Five agents, each with a distinct role, working in clockwise sequence to automate the entire hiring process end-to-end.

1

Resume Intelligence

Parses PDF resumes via GPT-4o-mini into structured JSON - extracting skills, work history, contact info, and achievements for downstream agents.

LLM Extraction
2

Interview Planning

Generates a 4–5 topic interview roadmap tailored to the candidate's profile and target role, covering technical depth, behavioral fit, and experience validation.

Adaptive Planning
3

Voice Interview

Places real outbound calls via Twilio with AWS Polly Neural TTS. Handles natural conversation, barge-in interruptions, and silence recovery in real time.

Live Telephony
4

Decision & Control

State-machine orchestrator managing the full call flow - identity checks, skip logic, follow-up retries, silence recovery, and real-time prompt injection detection.

Orchestration
5

Evaluation & Scoring

Post-call, scores the full transcript against the rubric - correctness, communication, professionalism, and job fit - producing a detailed breakdown and final Accept / Reject verdict.

Rubric Scoring

Core Capabilities

What makes HireLens more than just a chatbot.

Barge-In Support

Handles natural interruptions mid-speech using Twilio's real-time barge-in - no robotic "please wait" loops. Silence is gracefully recovered with prompts or topic escalation.

Tamper Detection

Detects prompt injection, jailbreak attempts, and candidates trying to game the AI in real time. The session is immediately flagged and terminated with a Reject verdict.

Adaptive Follow-Ups

Vague or weak answers trigger targeted follow-up questions - up to 2 attempts per topic. The agent logs the gap, then moves on without letting a single question derail the session.

Automated Verdict

A custom rubric is generated per candidate before the call. Post-call scoring produces a per-item breakdown across correctness, communication, and job fit - outputting a clean Accept / Reject.