Skip to content

Second Chair

A voice agent built for three people in the room rather than two on a call. It tracks who is speaking, warns the professional privately in their earpiece when a rule is crossed, and stays silent the rest of the time.

Open second-chair-chi.vercel.app

Product design, real-time audio pipeline and full-stack build

3-party

Built for a room, not a two-person call

Dual channel

Private warning in the earpiece, public answer in the room

Cited

Every warning names the rule that triggered it

What made it hard

Every voice agent on the market assumes two participants: a user and an assistant. A compliance conversation has three — the professional, the client, and whoever came with them. That breaks the usual design in three places at once: the agent has to know who is speaking, it has to warn one person without the other hearing, and it has to stay quiet through a conversation it is not part of.

How it was built

  1. 01

    Knowing who is speaking

    Speaker diarization in real time, because a compliance rule applies to the professional and not to the client. Attributing a sentence to the wrong person turns a useful warning into noise, and noise is what gets an assistant switched off.

  2. 02

    Two channels, two audiences

    Guidance is whispered to the professional on a separate audio path while answers to the room go out loud. The dual-channel design is the product: a warning the client can hear is worse than no warning at all.

  3. 03

    A deterministic tier in front of the model

    The rules that must fire immediately are evaluated deterministically rather than waiting on a language model. Model-based evaluation handles the judgement calls; anything that has to be instant never queues behind a token stream.

  4. 04

    Silence as a designed behaviour

    It answers when addressed by name and otherwise says nothing. Most of the engineering in a live-listening agent goes into deciding when not to speak, which is the opposite of how these systems are usually demoed.

  5. 05

    Redaction before storage

    Account numbers and identifying details are stripped before a transcript is written. The audit record has to survive a compliance review of its own, so the safest transcript is the one that never held the number.

What it does

  • Real-time speaker identification
  • Private earpiece warnings
  • Answers only when addressed
  • Compliance rule evaluation
  • PII redaction before storage
  • Cited audit transcript
More work

Other things we've shipped.

AI compliance platform

Sovereign

A self-hosted platform that audits contracts against GDPR, HIPAA and SOC 2 and returns cited, severity-ranked risk reports — with inference running on the customer's own GPUs, so no document leaves their network.

  • React
  • RAG / vector retrieval
  • LLM orchestration
  • Self-hosted inference

AI productivity platform

Montlify

One workspace holding 33+ AI tools for writing, research and chat, with the accounts, subscriptions and infrastructure needed to keep them all running.

  • React
  • LLM APIs
  • Streaming
  • Authentication

Have something like this in mind?

A short call, then a written scope with a fixed price and a date. Scoping costs nothing and the document is yours either way.

Scope a project

Replies within one business day