← All speakers

Bio, Work & Ideas

Ornella Bahidika

Conference affiliation: Microsoft · 2026

Ornella Bahidika is a Microsoft Teams product manager and co-creator of Ace, a voice-first AI tutor. Her work addresses a practical challenge in AI products: making conversational systems responsive and reliable without allowing language models to control essential application logic.

Originally from Bukavu in the Democratic Republic of the Congo, Bahidika studied business administration and information technology management at Georgia Tech, where she served as an undergraduate sustainability ambassador and graduated in 2022. At Microsoft, she has worked on Teams and data-driven workplace collaboration, highlighting Cloud IntelliFrame occupancy reporting for understanding how meeting spaces are used.

With Joel Allou, Bahidika built Ace to deliver complete spoken lessons without skipping stages, repeating itself, or stopping prematurely. Its architecture reflects two distinctive engineering priorities:

  • State-machine orchestration: Application code controls the introduction, instruction, comprehension checks, grading, progression, and conclusion. Each model interaction receives a limited assignment; the surrounding software validates its response, determines the next step, and governs features such as the shared whiteboard. Their agent-control demonstration illustrates how conversational fluency and operational authority remain separate.
  • Latency-driven model selection: Ace targets approximately 950-millisecond voice-agent latency, making response speed central to the tutoring experience. Rather than requiring a large model to reconstruct lesson plans and student context on every turn, deterministic model scaffolding prepares that information externally. Bahidika and Allou demonstrated Claude Haiku 4.5 responding in roughly 900 milliseconds, compared with a slower Claude Opus 4.7 configuration. Smaller models still require structural guardrails against drift, but explicit orchestration can make them practical for real-time interaction.

Read the topics behind these talks

2 conference talks

Key ideas

Scroll to read ↓

Ace’s live voice tutor puts lesson progression in a state machine, leaving the model to perform bounded actions while a harness validates results and decides what happens next.

  • The lesson ends halfway through
    0:00 ↗
  • Give each step a narrow contract
    0:53 ↗
  • Smaller assignments enable a smaller model
    1:24 ↗
  • Control the operations around the conversation
    3:35 ↗
  • The model proposes; the harness decides
    4:21 ↗
  • Move unreliable control flow into code
    5:02 ↗

Key ideas

Scroll to read ↓

Ace keeps its voice tutor responsive by moving lesson state, mastery tracking, and next-step decisions into code, leaving a small model to turn those decisions into speech.

  • The pause is the problem
    0:00 ↗
  • Give the model a turn summary, not the whole lesson
    1:04 ↗
  • The same question, two different systems
    3:07 ↗
  • Pay for structure in code
    4:15 ↗
  • A small model inside a larger system
    5:16 ↗

References