← Back to digest
Tools & Demos

Whodunnit: voice-driven murder mystery with AI suspects

What's this about?

Whodunnit is a Show HN game where you solve a murder mystery by *talking out loud* to AI-played suspects, interrogating them and catching them in contradictions. It's a fun, polished demo of real-time voice agents that maintain a consistent story.

What exactly is a "voice agent"?

A voice agent is an AI you interact with by speaking — it transcribes your speech, reasons about it, and replies in a generated voice, in near real time. The hard parts are low latency (so conversation feels natural) and memory/consistency (each suspect must stick to their alibi across the whole interrogation). Whodunnit turns exactly those hard parts into the gameplay.

What's the community saying?

HN 192 pts, 81 comments
  1. Genuinely fun — commenters found the interrogation loop engaging, a rare "AI demo I actually enjoyed" reaction.
  2. Latency and voice quality — discussion of how snappy the back-and-forth feels and which voice stack makes real-time interrogation work.
  3. Consistency is the trick — interest in how the suspects keep their stories straight (and whether you can break them with clever questioning), i.e. the memory problem made playable.
← Macaron-V1: open continual learning with mixture-of-LoRA