Tough Round AI

A mock interview only helps if it feels real, which means low-latency speech in both directions, an interviewer that adapts to the answer it just heard, and feedback specific enough to act on. No single AI provider is best at all of that, and candidates wanted a desktop app that behaved like a native tool rather than a browser tab.

We built one React + Vite front end shipped as both a web app and a packaged desktop app, over a Nest.js API that orchestrates several providers behind a single internal contract — Gemini, OpenAI, and Anthropic for reasoning and feedback, Deepgram for speech, and Google Cloud for supporting services. Sessions, transcripts, and scoring live in Neon Postgres through TypeORM; Redux holds live session state so a dropped connection does not lose an interview in progress.

Candidates run spoken mock interviews that adapt as they answer and get structured feedback afterwards, on desktop or in a browser from the same codebase. Because model access sits behind one internal boundary, providers can be swapped or added without touching the interview flow.