Viva

01The problem

Help students explain code they shipped with an AI agent's help, before an interviewer asks about it.

Students now ship projects an AI agent half-wrote, and interviewers know it. Generic mock interviews don't ask about the code you actually have.

02What I built

A Mastra agent on Gemma, an open-weight model, with model output checked in code: file paths against the real file list, follow-up quotes against the answer, and scores computed outside the model.

• Reads a public GitHub repo (or a local folder) and writes a short brief; invented file paths are dropped
• Uses Entire checkpoints to find which files an AI agent wrote, and the prompt that started it
• Runs a five-stage interview; the agent opens files itself with a readFile tool
• Every follow-up must quote the candidate's answer, and code checks the quote is real
• Scores each answer 0 to 4 in code; Gemma writes the coaching text
• Local mode remembers weak spots across sessions and can run fully offline on Ollama

03Why it won

Best Use of Entire at DEV's Hacktoberfest 2026: it reads an agent's checkpoints to find the files the AI wrote, then interviews you on exactly those.

04Results

Interview stages

5

Score scale

0–4

Modes

Hosted + local

05Stack

• Next.js
• React
• TypeScript
• Mastra
• Gemma
• Ollama
• LibSQL
• Vitest