Lumina

Lumina is an AI teaching system designed to deliver personalized, interactive, and responsive real time learning. Instead of acting as a chatbot that simply answers questions, Lumina behaves like a real teacher. It plans lessons, explains concepts, checks understanding throughout the lesson, adapts when a student is struggling, and only moves forward once the student has demonstrated understanding. The proposed solution doesn't add more educational content to a world already full of it, it turns content that already exists (a teacher's curriculum, a textbook, a set of lecture notes) into a live, responsive teaching process. Two tiers, both running on AMD Instinct hardware: Fast tier; Qwen3-14B, self-hosted on our own AMD Instinct GPU instance via vLLM. This is the tier answer engine calls on every single learner turn checking an answer, grading a response to a probe, deciding whether to escalate to a different teaching approach. It must return in a few hundred milliseconds or the "live teacher" illusion breaks Reasoning tier; Qwen 3.7 Plus, hosted on Fireworks AI's AMD Instinct MI300X infrastructure. This handles lesson planning and the harder in-lesson explanations, where we deliberately let the model think longer using Fireworks' reasoning effort parameter, a tuneable control exposed by Qwen 3.7 Plus's hybrid-reasoning mode. Planning happens once per lesson, so the extra latency is worth spending on quality output. Both tiers deliberately run the same model family, Qwen's hybrid-reasoning checkpoints rather than mixing model families across tiers. That was an intentional design decision, not a coincidence a lesson planned by one model family and taught turn-by-turn by a different one risks a mismatch in how the two explain and phrase things. Keeping one model family across both tiers, differing only in size and in where each instance runs, is what keeps that inconsistency from ever surfacing in the actual lesson.