Joel Allou

Joel Allou co-created Ace, a voice-first AI tutor that uses explicit software controls to make interactive lessons faster and more reliable. Working with Ornella Bahidika, he designed an architecture that assigns conversation to a smaller language model while keeping lesson progression, student assessment, and application behavior under programmatic control.

At AI Engineer World’s Fair 2026, Allou demonstrated how this division of labor addresses two persistent problems in production AI: delayed responses and unreliable multistep execution.

• Voice-agent latency shapes model choice. Ace targets responses within roughly 950 milliseconds. In a comparison of voice-agent architectures, a scaffolded Claude Haiku 4.5 configuration responded in approximately 900 milliseconds, while a larger Claude model took longer. Allou argues that the most capable reasoning model can be the wrong choice when spoken interactions demand immediate feedback.

• A lesson state machine preserves instructional continuity. Application code tracks each lesson’s stage, prepares relevant context, and directs the model toward a narrowly defined conversational task. This deterministic scaffolding helps prevent the drift that smaller models can exhibit during extended interactions.

• Agent harnesses control consequential decisions. Ace’s orchestration layer validates outputs, manages whiteboard interactions, determines whether a student understands the material, and decides when to advance or end a lesson. Allou’s application-controlled agent orchestration applies the same principle to other multistep systems, including coding agents, onboarding flows, and operational runbooks.

Allou’s central engineering position is concrete: reserve language generation for the model, and implement control flow and output validation in software.