Practical Architecture for an AI Learning Platform

Building Sikho.ai over the last year taught me a lot about what an AI-learning platform actually needs architecturally. Sharing the highlights for other teams building in this space.

The data layer

Three stores:

1. Session state — conversation buffer, scratchpad, ephemeral context. Redis.
2. Learner state — mastery per concept, goals, preferences, long-tail history. Postgres.
3. Content state — lessons, problems, canonical explanations, curriculum graph. Postgres with pgvector for semantic search.

Prompts stitch across all three. Getting this right on day one pays dividends forever.

The inference layer

A tiny router in front of model calls:

• Fast path: cached response if the learner and the question match a prior interaction.
• Small model path: Haiku or similar for easy explanations and quick reactions.
• Big model path: full reasoning for hard concepts or struggling learners.
• Fallback path: graceful degradation if the primary model is slow.

The evaluation layer

Every interaction gets logged with: learner state snapshot, prompt, response, response time, learner feedback (thumbs/engagement). Weekly human review samples interactions for quality.

This is the part most teams skimp on. It is also the difference between a product that works and one that silently regresses.

The guardrails layer

Fact-checking on high-stakes claims. Confidence thresholds that trigger "I'm not sure" responses. Safety classifiers for age-appropriate content. These run post-model, pre-render.

The UX layer

Streaming responses. Persistent sidebar for history. Clear visual signals when the tutor is thinking vs. uncertain vs. confident. Zero magic — learners need to understand what the AI is doing and why.

Where we go next

Every layer is still evolving. Memory is the next frontier — longer-term, lighter-weight, smarter retrieval. At Sikho.ai we are building for learners that the system will serve for years, not sessions. Follow our journey @sikhoverse on Instagram, YouTube, and Facebook.

If you are building an AI-learning product, come compare notes. There is room for many winners and the playbook is still being written.