ColdMock: Building an Autonomous, Zero-Cost Voice Interviewer for My College Peer

This article reports on a project called ColdMock, which was created for the Hacktoberfest Weekend Challenge: Build for a Friend. The author, a college peer, built the tool to help his friend prepare for campus placements. While the friend excels at solving coding problems, he struggles with explaining systems concepts in real interviews because the AI tools he tried charge subscriptions or take too long to respond.

ColdMock addresses these issues by enabling the user to conduct rapid-fire oral technical drills on their laptop without any cost. The process works as follows: the AI reads out a realistic backend/system scenario, the user talks into their laptop microphone to explain their approach, and if there's a pause of 5 seconds, the AI automatically submits the answer.

The AI then generates a spontaneous follow-up question based on the user's previous response. At the end, it provides a scorecard detailing the user's strengths, blind spots, and a realistic hiring verdict.

The project runs entirely in the browser, with the AI model, Google Gemma 2 (2B), running offline via Ollama. The backend is a simple FastAPI service that manages the interview state and strict grading prompts, automatically marking blank or one-word answers as incorrect to prevent users from cheating. The frontend is built using vanilla HTML/JS with Tailwind CSS and the browser's built-in Web Speech API for both speaking and listening. There are no paid voice APIs required, making the tool free for students.

The open innovation aspect of ColdMock is significant, as it eliminates the need for students to pay for API fees or subscriptions to practice speaking through code. The tool ensures privacy by keeping all audio and interview notes on the user's laptop and operates offline, ensuring uninterrupted practice even in areas with spotty Wi-Fi connections.

The creator's friend tested the tool with a 3-question drill on database locks. When prompted about handling replication lag using read replicas, ColdMock challenged him on this aspect, a question he had previously struggled with in interviews.