Evolv: Self-Improving Behavioral Interviews

The Problem We're Solving

After spending 5 years in market research at firms like Kantar and IMRB, I've conducted hundreds of behavioral interviews. The dirty secret of qualitative research? We keep asking the same questions even when they don't work.

Some questions consistently yield generic, surface-level responses ("What's your greatest strength?"), while others unlock deep, specific insights ("Tell me about a time you failed..."). But we have no systematic way to identify which is which—until now.

What It Does

Evolv is a self-evolving interview agent that:

• Conducts live behavioral interviews via voice (VAPI)
• Scores each response in real-time across 4 dimensions: Specificity, Depth, Behavioral Evidence, and Novelty
• Automatically re-ranks questions based on performance
• Shows which questions are worth asking—and which to retire

During a 3-minute interview, you'll see questions jump from position #8 to #3 based on the quality of insights they produce. It's meta-research: using AI to optimize research itself.

Inspiration

The self-evolving agents hackathon theme hit me immediately: what if the questions themselves could evolve? Traditional A/B testing requires hundreds of interviews. With AI scoring, we can identify high-performing questions in real-time, within a single conversation.

This bridges my two backgrounds—research methodology and AI engineering—to solve a problem I lived with for years.

How We Built It

Architecture:

• Voice Layer: VAPI handles the phone call and transcription
• Scoring Engine: LLM (GPT-4) analyzes each response and returns structured scores
• Frontend: Next.js dashboard built with Lovable, showing live question rankings
• Billing: Fastino tracks usage-based API costs in real-time
• Validation: Tavily validates response novelty against web data

Technical Flow:

Built With

• anthropic-claude-(development-assistant)-**sponsor-technologies:**-fastino-(usage-based-billing-tracking)
• company-context)-**real-time-communication:**-websockets-/-server-sent-events-**deployment:**-vercel-**development-tools:**-node.js
• fastino-api
• framer-motion-**voice-&-ai:**-vapi-(voice-agent-platform)
• google-deepmind-api
• google-deepmind-gemini-(llm-scoring-alternative)
• javascript-**frameworks:**-next.js-14
• languages:**-typescript
• linkup-(job-&-company-data-enrichment)-**apis:**-vapi-api
• linkup-api-**data-enrichment:**-linkup-(candidate-job-history
• lovable-(ui-development-&-rapid-prototyping)
• openai-api
• openai-gpt-4-(response-scoring)
• react
• tailwind-css
• tavily-(web-search-&-novelty-validation)
• tavily-api