An AI-first product reached an $8.25m valuation.
In four years, with zero room for hallucinations.

In 2021, everyone else was bolting AI onto the same old product.

The EdTech market was crowded with content-heavy platforms competing on breadth rather than depth. Most existing products in 2021 weren’t really thinking about AI at all – it was tacked on, not built in. To win customer preference and justify the switching cost, this product had to be better than the competition on quality alone.

The catch: generative models hallucinate, and a wrong or pedagogically unsound answer is unacceptable in a learning context. Ensuring accuracy and eliminating hallucinations wasn’t a nice-to-have feature – it was the line between a product teachers could trust and one they couldn’t.

Three constraints were non-negotiable:

• Win a crowded, content-heavy market
• Be better than the competition, not just different
• Eliminate hallucinations – zero tolerance for unsound feedback

AI as the product, not a feature bolted onto it.

The team built an AI-first adaptive tutor where personalisation and instant feedback were the core value proposition, not an add-on. A learner model combined performance history, curriculum mapping and engagement signals to generate adaptive lesson sequences, with reinforcement learning and supervised models prioritising the exercises that closed each student’s knowledge gaps.

Every explanation, hint and graded response was generated in real time – and grounded, not guessed. Retrieval-augmented generation tied outputs to a vetted knowledge base and curriculum-aligned content, while fine-tuning on domain-specific datasets improved fidelity further. A validation layer cross-checked AI outputs against canonical answers and confidence scores: low-confidence responses fell back to a hint or an ‘ask a teacher’ prompt, or were routed to human review. Speculative reasoning and policy-related content were explicitly disallowed and filtered at inference.

Continuous A/B testing and human annotation teams audited model responses and fed corrections back into fine-tuning, while explainability features showed teachers why a recommendation was made, so they could trust it or override it. On top of that, the product-led mechanics did the rest: a quick diagnostic test, a personalised study plan and a visible score improvement at first touch, plus shareable progress reports and teacher collaboration tools that turned users into referrals.

Four years, one AI-first product, an $8.25m valuation.

The team attributes the outcome directly to AI-led product quality: they don’t believe the rapid scale and differentiation would have been possible without AI that delivered precise, trustworthy, pedagogically sound interactions. The discipline of eliminating hallucinations and building conservative guardrails is what let the product beat competitors who treated AI as an afterthought.