Miju Labs

← All people · Centaur Labs (trading as Centaur AI)

MIT-trained collective-intelligence researcher who turned a thesis finding (crowds of medical students can out-classify individual dermatologists) into Centaur, the best-known gamified expert-annotation company in medicine. His core idea is performance-weighted expert crowds rather than credential-weighted ones.

high confidenceBostonAt Centaur Labs (trading as Centaur AI) since 2017Updated 2026-09-19

Background

Duhaime did his PhD at MIT Sloan's Center for Collective Intelligence under Thomas Malone (completed 2019), after an MPhil in human evolution at Cambridge and a BA in economics and biology at Brown (per YesPress profile).

The founding story, as told in profiles: he watched his wife, then a medical student, spend hours on quiz apps, and combined that with research showing aggregated student judgements beat individual specialists on skin-lesion classification. The result was DiagnosUs, a competitive diagnosis app that pays for accurate answers and emits labeled data. Centaur went through Y Combinator (W19) and raised a $15M Series A led by Matrix Partners (2021) and a $16M Series B (2024).

What they run now

Runs the company and is its main public voice. Current messaging (2026) is about dynamic labeling pipelines: aggregate several diverse experts, use disagreement to escalate hard cases, and deliver confidence estimates with labels. The website shows a push beyond healthcare into LLM/software, insurance, robotics and finance customers.

Career

  1. 2017 – presentCo-founder & CEO, Centaur Labs / Centaur AI
  2. 2019PhD researcher, MIT Center for Collective Intelligence
  3. completed 2019MIT Sloan School of Management, PhDcollective intelligence, advisor-lineage with Thomas Malone
  4. University of Cambridge, MPhil (Human Evolution)
  5. Brown University, BA (Economics & Biology)

On the record

Credentials vs performance

Expert opinions should be weighted by measured accuracy, not years of experience or credentials. source

Diversity of annotators · 2026-03

You want more than one annotator per case, and not several who all think alike. source

Disagreement as signal · 2026-03

When annotators disagree, the case should be escalated automatically for more votes; finding controversial cases early is valuable. source

Human + AI aggregation

Aggregating humans and models beats either alone (Centaur cites 92.1% vs 91.4% top human vs 90.8% top AI on a skin-lesion task). source

Recent posts

1 posts archived

Zach Rausnitz reposted this
Erik Duhaime
Cofounder/CEO at Centaur.ai
A classic collective intelligence example (one that all MBA students will have done in one class or another!) is asking people to guess how many jelly beans are in a jar, then showing that the average opinion is often the best estimate. That also works when it comes to estimating how many calories are in a dish based on a photo -- and for many more consequential things, like determining whether there is cancer in a photo of a skin lesion or in a CT scan. And that's the case even when the best AI models are better than the best people. Collective Intelligence > Artificial Intelligence
471 comments12 reposts

Interviews & talks

Connections

  • Zach Rausnitzco-founder
  • Tom Gellatlyco-founder
  • Thomas MaloneMIT CCI founding director; Centaur advisor
  • Alexandr Wangindividual investor in Centaur (listed on about page)

Who they amplify

Accounts whose posts Erik has reposted recently: Centaur.ai.

Why it matters here

Duhaime has the most mature public playbook for the problem Julian faces in taste data: how to judge expert quality when credentials don't predict accuracy. Performance-weighted aggregation, disagreement-driven escalation and a gamified status economy to recruit supply all transfer directly to visual/design judgement, where 'taste' is also hard to certify. Direct market overlap is low (clinical, not creative).

How to reach

Speculative: he writes and podcasts about collective-intelligence methods, so a concrete question on applying performance-weighted crowds to subjective (non-ground-truth) domains like aesthetics is the most likely hook. LinkedIn is the verified channel; MIT CCI network is a possible warm path.

What we could not establish
  • Personal X handle not verified (only company @centaurlabs found).
  • YC profile lists a different LinkedIn URL (linkedin.com/in/erik-duhaime-7a827413/); which is canonical is unclear.
  • No 2026 funding, revenue or frontier-lab customer information found.
  • Dates of the podcast appearances not verified.

Sources

  1. Y Combinator: Centaur
  2. YesPress: Erik Duhaime profile
  3. Centaur.ai about page
  4. Centaur blog: From Prediction Markets to Production-Ready Data (2026-03-05)
  5. Centaur author page: Erik Duhaime
  6. Centaur.ai homepage
  7. MIT News: Gamifying medical data labeling (2023)
  8. LinkedIn: Erik Duhaime