The green owl mascot in reading glasses, standing and looking around

AI productIntervooCase study

Intervoo

Live AI interview copilot.

An AI product that helps candidates practice and perform in real interviews — with live voice and a privacy-first design.

Role
Founder & Principal Engineer
When
2024 to now
Team
Solo build
Read
About 4 min
The green owl mascot in reading glasses, standing and looking around

Overview

Overview

Intervoo is a live AI interview copilot.

It listens to interviews in real time, recognizes when a question is asked, and quietly suggests a personalized answer based on the candidate's resume and the company they're interviewing with. A separate practice mode helps candidates prep beforehand and quietly builds a personal story library that makes future copilot sessions more accurate. Designed, built, and shipped end-to-end.

The challenge

A live interview is the worst place to put an AI tool.

Latency has to be under a second or the conversation moves on. Answers have to actually sound like the candidate, not a generic chatbot. And privacy is non-negotiable — nobody wants their interview audio sitting on someone else's server. Most existing tools fail one or more of these. The hard part wasn't the AI — it was building the whole experience around those three constraints.

  • Under a secondor the moment passes
  • Sounds like younot like a chatbot
  • Private by designno stored audio

The solution

Two coupled modes on one design.

Two coupled experiences sharing one design: a practice mode that helps candidates prepare and quietly extracts their best stories, and a live copilot mode that uses those stories — plus the target company's culture — to generate personalized suggestions in real time. Voice processing happens at the edge, audio is never stored, and the entire system is designed to disappear behind the conversation rather than interrupt it.

Architecture

Three real-time loops.

Three real-time loops: voice in, AI out, and a practice flywheel that quietly improves both. Privacy is enforced at the system boundary, not bolted on after.

Client
  • Browser-side voice capture
  • Streaming UI
  • Question detection
Voice
  • Real-time transcription
  • Speaker separation
  • Zero audio retention
AI
  • Streaming LLM responses
  • Personal context grounding
  • Company culture awareness
Data
  • Encrypted user data
  • Story library
  • Privacy-first storage

One request, start to finish

  1. Candidate speaks
  2. System recognizes a question
  3. Personal context is assembled silently
  4. AI suggestion streams to the candidate
  5. Practice sessions quietly improve future suggestions

Decisions

Six calls and why I made them.

Call 01

Streaming, not request/response

Sub-second perceived latency was non-negotiable. Streaming the response token-by-token lets the candidate start reading almost immediately, instead of waiting for a full answer to load.

Call 02

One model, hand-written orchestration

Tried frameworks early on. Every layer added latency, complexity, and cost without earning its keep. Stripping it back to one streaming endpoint with custom prompt logic was faster, simpler, and far easier to debug.

Call 03

Practice as the data engine

Practice answers are stories the user has already validated. Quietly extracting them into a personal library makes future copilot answers feel personal without forcing the user to fill out a profile. Two features, one shared flywheel.

Call 04

Zero audio storage, ever

A live interview tool that uploads audio is a non-starter for trust. Audio is processed in the browser and discarded. Only what the user explicitly saves persists.

Call 05

Resume as single source of truth

Earlier iterations fragmented user data across multiple tables. Consolidating to one source of truth eliminated entire classes of sync bugs and made personalization straightforward.

Call 06

Database-level access control from day one

Multi-tenant data with AI in the loop is a leaking-data accident waiting to happen. Enforcing isolation at the database level means a bug in any API route can't cross user boundaries.

Reflections

What I took away.

Frameworks cost more than they save on latency-critical paths

I started Intervoo on a popular LLM orchestration framework because it seemed irresponsible not to. Three weeks in, every prompt I wanted to tune meant fighting an abstraction, and the streaming was worse than a raw fetch. Ripped it out over a weekend, dropped to one streaming endpoint with hand-written prompt logic, and perceived latency dropped noticeably. The lesson that keeps coming back: on the hot path, frameworks are renting complexity you'll want to own later anyway.

Privacy constraints make the product, not break it

"Zero audio storage" sounded like a limitation at first. It turned out to be the single clearest positioning we had. Once I committed to it, a lot of architectural questions answered themselves — processing at the edge, ephemeral sessions, no recording features. Constraints at the start of a project are a gift if you treat them as axioms instead of problems.

Practice and copilot shouldn't be separate features

The first version of Intervoo had two product surfaces that didn't talk to each other. Users would practice, and then show up to the live copilot with no personal context. Coupling them through a shared story library — where practice quietly builds the data the copilot uses — is what turned two decent features into one compounding product.

Outcomes

What shipped.

  • LatencySub-secondTime to first token, streamed to the candidate
  • AudioZero audioStored anywhere, ever
  • ModesTwo modesPractice and live copilot on one data model
  • BuildSoloProduct, AI, full stack, design and deployment

Stack

Technology

Frontend

  • Next.js
  • React
  • TypeScript
  • Tailwind CSS

AI & Voice

  • Streaming LLMs
  • Real-time STT
  • WebRTC

Backend

  • Supabase
  • PostgreSQL
  • Edge Functions

Infra

  • Stripe billing
  • Netlify

Features

What it does

  • Streaming LLM pipeline that delivers token-by-token responses with sub-second time-to-first-token
  • Real-time voice layer: browser-side audio capture, low-latency transcription, live question detection
  • Personalization engine that grounds every suggestion in the candidate's own story library and target company context
  • Privacy-first architecture: audio processed at the edge and never persisted — only transcripts the user explicitly saves
  • Two coupled product surfaces sharing one data model: practice mode quietly extracts stories, live mode uses them
  • Multi-tenant Postgres with Row Level Security — a bug in any API route can't cross user boundaries
  • Stripe billing, webhooks, and subscription state wired into a Next.js full-stack on Supabase
  • Solo build: product, architecture, AI engineering, full-stack, design, infra, deployment

Keep reading

More from the AI Lab

Your project

Building something like this?

AI, voice, privacy constraints or a mobile app that can’t afford to break. Free 30-minute call, no pitch, clear next steps either way.

Email
Back to
All work