Areté - AI Technical Interview Platform

An AI interviewer that conducts LeetCode-style technical interviews via voice call, watches code in real-time, provides live hints, detects hiring bias, and generates comprehensive recruiter reports.

Inspiration

Technical interviews at scale are painfully expensive and inconsistent. Companies pay $200/hour for senior engineers to conduct repetitive LeetCode-style interviews, time that could be spent building products. Worse, human interviewers introduce unconscious bias, inconsistent question difficulty, and varying scoring standards. We asked: What if AI could conduct fair, consistent, and cost-effective technical interviews 24/7?

The spark came from realizing that real-time code analysis combined with conversational AI could create an interview experience that actually feels like talking to a human interviewer—one who watches your code, offers hints when you're stuck, and encourages you when you're on the right track.

What it does

ARETE is a multi-agent AI interviewer that:

Conducts live voice interviews using LiveKit's real-time voice infrastructure
Presents LeetCode problems (Two Sum, Merge Intervals, etc.) via natural voice conversation
Watches code in real-time through a Monaco editor, analyzing snapshots every 1.5 seconds
Provides intelligent hints when it detects logical errors (not typos—those it lets you fix)
Runs candidate code in a secure sandbox and evaluates against test cases
Scores performance across correctness, optimization, communication, and problem-solving
Detects hiring bias through a dedicated Fairness Monitor agent that analyzes transcripts
Generates recruiter dashboards with scores, recordings, transcripts, and bias reports

Cost: ~$1 per interview vs. $200 for a human interviewer → 99% savings.

How we built it

Backend Stack

FastAPI for REST and WebSocket APIs
LangGraph for multi-agent orchestration (Interviewer → Fairness Monitor pipeline)
Llama 3.3 70B on Groq - Powers the Interviewer agent with ultra-low latency inference
Gemini 3.0 Pro - Powers the Fairness Monitor agent for post-interview analysis
LiveKit Agents for real-time WebRTC voice communication
ElevenLabs for natural text-to-speech

Frontend Stack

Next.js 14 + React 18 for the web interface
LiveKit React SDK for voice call integration
Monaco Editor for the VS Code-style coding environment
TailwindCSS for styling

The 2-Agent System

Agent 1: Technical Interviewer - Conducts the live interview, monitors code in real-time, interrupts with hints on logical errors, encourages good approaches, and scores the solution
Agent 2: Fairness Monitor - Runs post-interview to analyze the transcript for bias, normalize scores against historical data, and flag any issues

Challenges we ran into

Natural conversation flow - Getting voice responses under 1 second to make conversations feel truly natural required extremely low-latency inference; Groq's Llama 3.3 70B hosting was critical here
Code analysis vs. linting - Teaching the agent to distinguish between typos (ignore) and logical errors (interrupt) without being annoying
Fairness quantification - Defining bias detection heuristics that are both meaningful and not overly sensitive

Accomplishments that we're proud of

Sub-second latency - Conversations feel natural and responsive, not robotic
Real-time code intelligence - The interviewer genuinely understands code evolving live, not just static analysis
Built-in fairness - Bias detection isn't an afterthought; it's a core agent in the pipeline
24-hour shipping - Two agents orchestrated, voice conversations flowing, code executing—all in one hackathon

What we learned

Multi-agent systems are powerful - Separating concerns (interviewing vs. fairness auditing) made both agents better at their specific jobs
Latency is everything for voice AI - The difference between 500ms and 2s response time is the difference between natural and awkward
Real-time AI is possible - With proper debouncing, smart prompting, and fast inference (Groq), sub-second responses are achievable

What's next

Live avatar support - Add realistic lip-synced video avatars for a more immersive interview experience
Multi-language support - Expand beyond Python to JavaScript, Java, C++
Resume parsing - Pre-analyze candidate backgrounds to tailor interview difficulty
Behavioral interview mode - Add a second interview type for soft-skill assessment
Company-specific training - Fine-tune on a company's past interview data and rubrics
Enterprise dashboard - Analytics across thousands of interviews, A/B testing different problem sets
Interview practice mode - Let candidates practice unlimited interviews before the real thing