Interview Copilot

Inspiration

Interviews reward the candidates who tell the clearest stories about their own work — but under pressure, most of us blank out, ramble, or undersell what we have actually done. I wanted to build something that could put your own experience back in your head at the exact moment you need it.

What it does

Interview Copilot listens to a live interview, detects when the interviewer asks a question, and generates a first-person answer grounded strictly in your CV — never invented, never generic.

For example, when the interviewer asks "What is your experience with microservices?", Copilot answers from your real work: Chetak ERP, 10K+ daily active users, OAuth 2.0/JWT authentication, REST APIs. A Deep Dive button produces a STAR-format answer for behavioral questions, using a larger model for richer storytelling.

The main workflow is:

1. The microphone captures the interviewer's audio in real time.
2. Speech recognition turns it into a live transcript.
3. A question detector identifies when a real interview question was asked (not just conversational filler).
4. NVIDIA Nemotron, running on Nebius Token Factory, generates a first-person answer grounded strictly in the candidate's CV.
5. For behavioral questions, Deep Dive produces a structured STAR-format answer with a larger model.

How we built it

I built Interview Copilot as a web application with a FastAPI backend and a lightweight frontend using the Web Speech API — no installs, it runs in the browser. I wanted to make sure the answers were trustworthy, not just fluent, so the model is constrained to the candidate's CV and cannot invent experience that isn't there.

Quick answers use a fast 30B model (Nemotron-3.5-Lightning) so the response arrives while the question is still fresh; deep dives use a 120B model (Nemotron-3-Super) for richer STAR-format answers. The whole pipeline runs on Nebius Token Factory.

Challenges we ran into

The genuinely hard parts were not the models. Telling a real interview question apart from conversational filler in a live transcript took real tuning — people say "right?", "you know?", and "so…" constantly. Keeping latency low enough that the answer arrives while the question is still fresh meant picking the right model for each job instead of using one model for everything. And the biggest one: preventing the model from hallucinating experience that isn't on the CV, which I solved by strictly grounding every answer in the candidate's CV context.

Accomplishments that we're proud of

Most importantly, I built this around a problem job-seekers actually face — freezing up in interviews — and it runs on real production infrastructure with real NVIDIA models on Nebius, not a demo script. It is useful outside of a hackathon setting: anyone can open the live demo, grant mic access, and practice.

What we learned

An AI assistant is only as trustworthy as the ground truth you give it. A good copilot should handle the complexity in the background — listening, detecting, grounding — instead of making the user deal with it.

What's next for Interview Copilot

Session history so candidates can review practice rounds, a coaching mode that scores delivery (STAR structure, filler words), and CV upload from the UI so anyone can use it with their own background. My goal is to make every candidate walk into an interview with their best stories ready.

Built With

• renderpythonfastapinvidia-nemotronnebius-token-factory