Mochi — Voice AI Screening Agent

Draft case study. Content to be expanded.

Summary

Mochi is an intelligent voice screening agent built at Maki People that autonomously conducts phone interviews with candidates in real time.

Context and problem

Screening large candidate pools by phone does not scale with human recruiters.

My role and ownership

Currently leading Mochi. Personally responsible for:

  • Real-time audio orchestration framework (LiveKit)
  • Multi-agent decision system (finite state machine)
  • Turn-management framework (interruptions, end-of-turn detection, flow)
  • Self-hosted LLM cluster powering in-call reasoning

Technical implementation

Architecture

  • Real-time audio streaming and processing on LiveKit
  • Multi-agent decisioning coordinated by a finite state machine
  • Turn management for multi-party conversation, interruption handling, and end-of-turn detection
  • Self-hosted LLM cluster for low-latency reasoning and dialogue

Stack

  • Language: Python
  • Real-time: custom frameworks on top of LiveKit
  • AI / ML: self-hosted LLM cluster

Key decisions and trade-offs

To be added.

Outcome and evidence

In production at Maki People. Detailed metrics to be added.

Links