← All speakers

Bio, Work & Ideas

Nik Caryotakis

Conference affiliation: SuperDial · 2025

Nik Caryotakis is an engineer at Overtone, the matchmaking company founded by Hinge creator Justin McLeod, who previously helped build production voice agents at SuperDial. His work addresses a central challenge in conversational AI: making automated conversations complete useful tasks reliably outside controlled demonstrations.

At Stanford, Caryotakis coauthored research applying neural networks to atrial fibrillation detection and helped develop R-Net Prime, a question-answering system incorporating character-level representations, self-attention, and pointer networks. He also played on Stanford’s 2019 national-championship water polo team.

At SuperDial, previously called SuperBill, he applied those interests to healthcare insurance-call automation: systems that receive requests through APIs, spreadsheets, and electronic health record integrations, navigate insurer phone trees, speak with representatives, and return structured answers. Human fallback, call auditing, office-hour tracking, and knowledge accumulated from previous calls helped make sensitive administrative work dependable.

His approach to production voice AI emphasizes four practical commitments:

  • Conversation design over voice realism: An agent’s value comes from completing domain-specific work and adapting to unexpected answers, not sounding impressively human. Reading scripts aloud can expose awkward interactions before deployment.
  • Composable voice infrastructure: Pipecat handles orchestration, TensorZero routes generative responses, and Deepgram provides speech recognition; replaceable components support model upgrades, provider fallbacks, and greater control over healthcare workflows.
  • End-to-end conversational testing: Recorded audio, simulated phone trees, pronunciation controls, and processor-level latency measurements test whether the entire interaction succeeds.
  • Accessible, accountable automation: Accent-related errors, misleadingly realistic voices, and unpredictable behavior demand human oversight and involvement from people with diverse perspectives.

Read the topics behind these talks

1 conference talk

Key ideas

Scroll to read ↓

A convincing voice is only the beginning: reliable healthcare calls depend on conversation design, precise spoken identifiers, observable infrastructure, and a path to completed answers.

  • From a chat agent to a reliable conversation
    0:54 ↗
  • The product is a completed set of answers
    2:21 ↗
  • A deliberately ordinary prior-authorization call
    4:24 ↗
  • What the voice AI engineer owns
    6:51 ↗
  • Design the conversation out loud
    9:58 ↗
  • Use orchestration that leaves room for the application
    12:37 ↗
  • Own the model interface and observe the work
    13:19 ↗
  • The text is not the delivered speech
    14:30 ↗
  • Names, latency, upgrade paths, and failures
    16:05 ↗
  • Test through the telephone
    17:26 ↗

References