viva-ai — a mock interviewer that talks back

The problem

Placement interviews are a skill you mostly get to practice inside actual interviews, where the cost of practising is the opportunity itself. A useful mock interviewer has to hold a real-time spoken conversation, understand the code you reference, and grade you on something more specific than "communicate better."

Approach

[Architecture: the voice loop (STT → reasoning → TTS) and its latency budget, how code-awareness is wired in, and how the grading rubric is structured and rendered.]

Results

[Pilot usage, grading agreement against human raters, end-to-end response latency.]

What broke

[Interruption handling? Accent robustness? Grading consistency across runs?]