SVC—006 ● LIVE VOICE AI / HR FEB 2026

Voice
Pre-Screen
Agent

Autonomous HR Decisioning

A fully autonomous voice agent that conducts real candidate interviews, evaluates responses in real time via GPT-4, and dispatches a binary hire decision (Recommended / Not Recommended) via secure webhook — no human in the loop. Multi-agent LLM architecture handles questioning, evaluation, and decisioning as separate concerns.

GPT-4
CORE MODEL
Voice
INTERFACE
Auto
DECISIONING
OpenAI SDKGPT-4ElevenLabs APIFastAPIPythonWebhooks
// UI PREVIEW — LIVE INTERVIEW SESSION CONSOLE
PRE-SCREENING SYSTEM Interview Console
Interview in progress
CANDIDATE QUEUE
PR
Priya Ramesh
Backend Engineer
● LIVE
JK
James Kim
ML Engineer
✓ DONE
AM
Aria Müller
Data Engineer
– QUEUED
TS
Tom Singh
DevOps
– QUEUED
LIVE TRANSCRIPT — PRIYA RAMESH
AGENT
Can you walk me through how you'd design a rate limiter for a distributed API?
CANDIDATE
I'd use a sliding window algorithm with Redis. Each service node increments a counter per client ID with a TTL equal to the window size.
AGENT
Good. What's the risk with Redis as a single point of coordination across nodes?
CANDIDATE
Network latency and potential split-brain if Redis goes down. I'd add a local fallback counter and use Redis Cluster for HA.
REAL-TIME SCORING
TECHNICAL DEPTH88
COMMUNICATION82
PROBLEM SOLVING85
HESITATION INDEXLOW
QUESTIONS LEFT2 / 5
RECOMMENDED
Webhook dispatch on completion
02 // ARCHITECTURE

The Pipeline

CANDIDATE
↔
ElevenLabs
STT / TTS
→
FastAPI
Orchestrator
→
GPT-4
Evaluator
→
Decision
Engine
→
Webhook
Dispatch
STEP 01
Initiate Call
FastAPI triggers the interview session. ElevenLabs opens a voice channel. Agent greets candidate with a generated introduction.
STEP 02
Voice Loop
Candidate speaks → ElevenLabs STT transcribes → GPT-4 generates next question based on prior answers. TTS voices the response back.
STEP 03
Live Evaluation
After each answer, GPT-4 Evaluator scores across 4 dimensions: technical depth, communication, problem solving, hesitation index.
STEP 04
Decision Engine
After N questions, all scores aggregated. Decision Model applies weighted threshold logic → Recommended or Not Recommended.
STEP 05
Webhook Dispatch
Decision + full transcript + per-dimension scores dispatched via POST webhook to HR system. No human review required.
03 // STACK

Built with

VOICE
ElevenLabs API
Handles both STT (speech-to-text) and TTS (text-to-speech). Streaming TTS keeps conversational latency under 600ms response time.
REASONING
GPT-4
Two GPT-4 instances: one as Interviewer (question generation) and one as Evaluator (answer scoring). Separated for clear role boundaries.
DECISIONING
Decision Model
Custom Python scoring model. Weighted average of 4 GPT-4 evaluation dimensions. Binary threshold tuned per job role requirements.
BACKEND
FastAPI + Python
Session orchestrator. Manages conversation state, queues follow-up questions, and coordinates between STT, LLM, and TTS services.
DISPATCH
Webhooks
Signed POST webhooks deliver the final decision payload — candidate ID, score breakdown, decision label, and full transcript — to the HR system.
SDK
OpenAI SDK
Function calling enforces structured JSON output from the Evaluator agent, ensuring consistent score schema for the Decision Model.
See the rest
of the deployments.
← ALL PROJECTS