Designed a 0→1 AI Tutor that actually scales.
I owned interaction design, voice UX, and the curriculum tooling system, end to end.
1:1 tutoring drives significantly better outcomes. But modern classrooms and ed-tech systems rely on fixed, linear instruction, failing to adapt to individual learners and creating a gap between time spent and true mastery.
Students receiving 1:1 tutoring perform ~2 standard deviations better, with the average student outperforming ~98% of their peers.
At ~$50–$250 per session, scaling this level of personalized instruction is infeasible.
So the real challenge isn't improving learning outcomes, it's making them scalable.
What if 1:1 tutoring could be made scalable?
Enabling real-time, back-and-forth learning, mirroring how students learn with human tutors.
Providing real-time, 1:1 guidance at the moment of confusion.
At a per-step level, rather than evaluated only at the end.
Are made scalable, without the cost constraints of human tutors.
The experience had to quickly feel approachable and responsive, especially in classrooms where hesitation, awkwardness, or lack of confidence can break engagement within seconds. During early classroom testing, a student said within 10 seconds: "It doesn't feel like it's actually listening to me." That single line drove every voice UX decision that followed.
I had to define how students speak, listen, interrupt, recover from mistakes, and stay oriented, all while keeping the interaction natural and low-friction.
Designing for a voice-first AI Tutor required rethinking how content is created, so it can drive real-time, conversational teaching instead of static delivery.
I replaced text-heavy static slides with this: a whiteboard-style interface using handwritten visuals that mimic exactly how a human tutor explains concepts step-by-step.
Added a simple progress-based interaction layer that advances on correct responses to reinforce momentum and participation.
Refined voice interaction with a "hold to speak" mechanism, making it easier and more natural for students to talk to the tutor.
Validated with ed-tech leaders through early demos of the core experience.
Voice-first interaction consistently stood out for its ease of use and intuitiveness.
~25% higher completion rates observed in controlled sessions, with students reporting higher engagement.
I bet on a gamified progress layer first. It didn't move engagement, so I cut it rather than iterate on it, and redirected the budget into interaction quality instead.
I deliberately left out mastery guardrails in this phase to validate free-form interaction first. It was the right call early on, but it meant students could skip ahead without demonstrating understanding, a gap I had to close in Phase 3.
Content creation was highly time-intensive (~104 hours for 30 minutes), exposing a scalability gap and leading to the creation of Studio. Case study here ↗
Increased flexibility introduced complexity, making it harder for content designers and highlighting the need for better tooling.
Gave students the ability to connect with their AI tutor through a personalized voice, increasing their sense of ownership and lowering engagement barriers.
Introduced a wider range of voice options, allowing schools to select the tutor voice that best matched their classroom culture and student preferences.
Redesigned the content creation workflow from manual authoring to a tool-assisted pipeline, significantly reducing production time per topic.
View Case Study ↗Personalised voice selection increased student engagement and session return rate.
Voice variety improved adoption across diverse classrooms, with teachers gaining meaningful control over the tutor's persona.
Content production time reduced significantly, enabling faster curriculum expansion across subjects.
Voice preference data required thoughtful opt-in UX to collect without friction.
Not all voices performed equally across subject types, requiring ongoing curation and quality control.
Tool-assisted production still required curriculum expertise, limiting full self-service content creation.
Added mastery-gated progression that prevents students from advancing without demonstrated understanding, enforcing mastery step-by-step.
Improved the consistency and quality of AI tutor responses through tighter curriculum alignment and iterative refinement based on live classroom feedback.
Introduced patterns students already recognize from classrooms (pacing cues, structured prompts) to lower cognitive friction and ease the transition.
Built real-time progress indicators into the student experience, helping learners track mastery trajectory and reducing anxiety about performance.
Co-developed instructional sequences with teachers in the loop, incorporating direct classroom observations into the AI tutor's pedagogical approach.
Guardrail-controlled progression significantly reduced skip-through behavior across sessions.
Teacher collaboration led to measurably higher instructional quality and classroom fit.
Progress visibility improved student confidence and reduced mid-session drop-off rates.
Guardrails occasionally frustrated students who found gated pacing too restrictive.
Teacher collaboration cycles added lead time to content updates and new topic rollouts.
Progress UI required careful calibration to motivate without generating performance anxiety.
Designed an automated student join/leave system that eliminated manual teacher intervention, reducing session setup time and enabling seamless classroom transitions.
Introduced a canvas-based spatial layout for tutoring sessions, enabling flexible concept explanation and multi-step problem solving beyond linear instruction.
Auto ingress reduced average session setup time from ~5 minutes to under 30 seconds.
Canvas layout improved complex problem-solving sessions, especially in multi-step STEM topics.
Teachers reported significantly reduced operational overhead during live classroom sessions.
Auto ingress exposed edge cases in low-bandwidth environments, requiring fallback mechanisms.
Canvas flexibility increased design complexity, requiring new content creation guidelines.
Greater automation reduced teacher visibility into session flow, surfacing a need for new oversight tooling.
Learners reached performance levels comparable to traditional classroom learning, driven primarily by the guardrail-controlled progression system redesigned in Phase 3, which prevented advancement without demonstrated mastery.
Moved from controlled environments to live classroom deployments, testing the system under real-world constraints and learner behaviors.
Expanded across multiple schools, grade levels, and subjects while maintaining consistent learning outcomes and engagement patterns.
Improved outcomes and system robustness helped secure and expand pilots across schools, increasing pipeline momentum and go-to-market traction.
Transitioned from static content delivery to adaptive, real-time tutoring, differentiating the product in a competitive EdTech landscape.
Combined pedagogy, real-time interaction, and system intelligence into a learning experience that's difficult to replicate. The AI-native content tooling I designed reduced topic production from ~104 hours to ~4 hours, compressing the moat further with every release.
Owned the end-to-end learner experience: defining interaction patterns and instructional flow, setting direction for content designers, and validating improvements through live classroom pilots.
Engineers (2–3): Built the AI Tutor system, enabling real-time voice interaction, adaptive tutoring, and scalable classroom deployment.
Product Manager / Founder: Drove product direction and led pilot rollout with partner schools.
Content Designers (contract, rotating): Produced curriculum content using the authoring system I designed; I set the pedagogical guidelines and reviewed output for quality.