Voice AI is not a product; it's a system - - Voice AI Space Conference NYC

Overview of Maki and Moshi

Maki is an intelligence layer that integrates with Applicant Tracking Systems (ATS) to manage hiring processes from end to end. Its voice AI agent, Moshi, serves as a candidate screener and assessor. Moshi can speak up to 30 languages, answer candidate questions about company culture, salary, and roles, and collect administrative and competency data. Recruiters can access a platform showing candidate profiles, eligibility criteria, and scores across more than 300 soft and hard skills, complete with auditable reasoning, transcripts, and video recordings.

The Challenges of Voice AI in Hiring

Hiring is a high-stakes environment where errors carry significant consequences, such as career impact, legal exposure, brand damage, and bias. This differs from low-stakes applications like ordering food. A fundamental tension exists between control and naturalness:

• Chat-style AI (often driven by a single-prompt LLM) feels natural and engaging but tends to drift from objectives and struggles to maintain structure over conversations lasting longer than five minutes.
• Scripted interviewers offer high control and reliable, defensible data collection but can feel robotic, leading to candidate disengagement.

Many standard voice AI platforms fail in production because they lack the customization required for the final 10% of the system. These platforms often rely on hidden wrappers that obscure visibility, create vendor dependency regarding pricing and model upgrades, and offer no observability into conversation states or scoring decisions.

Maki's Pipeline Architecture

Moshi is designed as a pipeline of specialized components working in a tight loop rather than a single model. The pipeline consists of five layers:

1. Speech-to-Text (STT): The foundational layer that must handle interruptions, overlapping speech, disfluencies (such as pauses and filler words), background noise, and regional accents.
2. Large Language Model (LLM): Extracts relevant signals, identifies key details, and interprets candidate intent.
3. Business Logic and Finite State Machine (FSM): Instead of a single prompt, Moshi uses nested prompts and a multi-agent system where different agents validate inputs, generate responses, and track progress. This ensures consistent data collection and structured, auditable interviews.
4. Evidence Extraction: Conversations are extracted first and then evaluated against fixed criteria developed by an internal science team to ensure unbiased scoring.
5. Text-to-Speech (TTS): Converts the system's responses back into spoken voice.

System Trust and Evaluation

To ensure reliability, Moshi undergoes a rigorous testing and release process:

• Pre-release: Benchmarking and local testing with clients to align on tone and quality.
• Pre-screening and Launch: Running LLM-to-LLM simulations with multiple personas to quantify completion and error rates.
• Post-launch: Generating a scorecard for every call alongside continuous monitoring, alerting, and fine-tuning.

This framework provides organizations with stable conversations at scale, comparable candidate evaluations, explainable failures, and minimized risk during scaling.