LLM Integration in EdTech: Lessons from Duolingo, Brainly, and SoloLearn

various · Education · 2023 · source Agentic: no · promotional bias: low · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Education technology and personalized learning

Duolingo, Brainly, and SoloLearn needed LLM features that improve learning rather than merely generate content, requiring personalization, pedagogy, curriculum grounding, teacher feedback, and cost-aware production operation.

Constraints:

• The source is a panel transcript with auto-caption noise, so claims are grounded in direct transcript spans and the ZenML summary where needed.
• Education systems must match learner vocabulary, curriculum exposure, and classroom-like explanations rather than simply produce correct answers.
• Brainly supports many school systems and curricula, so it cannot assume one owned curriculum.
• SoloLearn found that a generic assistant is not enough to become a core differentiated product.

Architecture

The panel describes a constrained, feedback-driven approach to LLM integration in EdTech. SoloLearn uses production AI assistance for code explanations and trains systems with teacher feedback. Brainly augments prompts with learner behavior, cohort similarity, content traversal, and curriculum context to generate teacher-like explanations. Duolingo uses LLMs for interactive language conversations, explanations, personalization, and curriculum generation, while acknowledging that teaching-quality content needs optimization beyond out-of-the-box generation.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  Learner_context["Learner context"] --> Prompt_augmentation_layer["Prompt augmentation layer"]
  Prompt_augmentation_layer --> LLM_explanation_generator["LLM explanation generator"]
  Teacher_feedback_loop["Teacher feedback loop"] --> LLM_explanation_generator
  LLM_explanation_generator --> SoloLearn_AI_assistant["SoloLearn AI assistant"]
  LLM_explanation_generator --> Brainly_curriculum_adaptation["Brainly curriculum adaptation"]
  LLM_explanation_generator --> Duolingo_AI_learning_features["Duolingo AI learning features"]
  SoloLearn_AI_assistant --> Learner_facing_features["Learner-facing features"]
  Brainly_curriculum_adaptation --> Learner_facing_features
  Duolingo_AI_learning_features --> Learner_facing_features

Components

Data flow

Learner questions, behavior patterns, cohort signals, curriculum/content context, teacher feedback, and product constraints are used to augment prompts. LLMs produce explanations, simplifications, detailed follow-ups, conversation practice, or code explanations. Human teacher feedback and product evaluation inform which features remain in production and how prompts are improved.

Why this architecture

The panelists converge on a constrained-learning architecture: use LLMs with context, feedback, and product-specific pedagogy because factual generation alone does not teach well or personalize like a teacher.

Models used: LLMs · retrieval augmented generation · AI assistant · prompt augmentation · teacher-feedback-trained system

Scale

• The three panel companies represent close to a billion learners.
  “between the three of you uh we're close to representing about a billion Learners”transcript, stated
• SoloLearn has over 30 million registered users.
  “over 30 million registered users around the globe mostly organic growth”transcript, stated
• Duolingo had used AI for personalization for about a decade.
  “we've been using AI since close to the beginning of dual lingo like a decade ago”transcript, stated

Alternatives considered (roads not taken)

Build a generic AI assistant

Rejected because: SoloLearn found that another assistant might help but would not differentiate or become a core product feature.

Generate factual educational content directly

Rejected because: The panel says factual content generation is easier than generating content that teaches well.

Use RAG before simpler constrained features are proven

Rejected because: Brainly describes starting with simpler constrained features before later moving to RAG use cases.

Patterns

rag-retrieval (retrieval)

Brainly moved into retrieval-augmented generation only after constrained context-editing features were working.

human-in-the-loop-gate (evaluation)

Teacher feedback is used to train and shape systems because LLMs do not automatically produce pedagogically effective content.

eval-driven-development (evaluation)

The companies discontinue AI experiments that do not work and optimize teaching-quality features with feedback.

Tradeoffs

Start with constrained explanation features before RAG.

Chose Simplify/explain-more features with known context over Immediate broad retrieval-augmented generation.

Rationale: Brainly first used contexts and facts it was comfortable with, then expanded into RAG later.

Use teacher feedback to shape generated learning content.

Chose Human educator feedback loop over Pure model-generated educational content.

Rationale: The companies need teaching-quality output, not just correct facts.

Gotchas (what bit them)

Factual generation does not equal good teaching.

Root cause: LLMs can produce information but do not automatically structure it pedagogically.

Resolution: Use teacher feedback and product-specific learning-science constraints to train and shape generated content.

Personalized education needs curriculum and learner context.

Root cause: A useful explanation depends on what the learner has already been taught and which vocabulary their class uses.

Resolution: Infer cohort, content traversal, and learner context, then augment prompts before generation.

Generic assistants are not enough for durable product differentiation.

Root cause: A basic assistant may be useful but not central to the learning product's core value.

Resolution: Focus AI work on the company's teaching strength, such as making hard concepts easy to understand.

Evaluation

The panel describes production retention/discontinuation of features, teacher feedback, user adoption, and ongoing optimization of personalization and explanation quality rather than a single benchmark.

Metrics: feature usage · teacher feedback · learner personalization quality · production retention or discontinuation

Outcomes

• SoloLearn had a production AI assistant for explaining coding errors in human language.
• Brainly used context and prompt augmentation to move toward teacher-like explanations.
• Duolingo applied AI to personalization, interactive language conversations, wrong-answer explanations, and curriculum generation.
• The panel consensus was that LLMs need constraints and feedback to teach well.

The panel represents close to one billion learners; SoloLearn alone reports over 30 million registered users.