How it works
How Annelia
thinks.
No magic. A few well-chosen algorithms doing their jobs.
§1 · The mastery model
A probability, not a percentage.
Each student × concept carries its own evolving distribution.
Posterior
p(mastery | evidence) ∝ Beta(α + s, β + f) s = positive signal weight f = negative signal weight Signal weight is bounded — a single answer can't claim more certainty than the system can earn.
§2 · Post-turn pipeline
Five stages. Always async.
Never blocks the next message.
- 01
Concept extraction
What did the student engage with? Not what's in the document — what got touched.
- 02
Confidence classifier
Confident, hesitant, or confused? Tone is signal.
- 03
Bayesian update
Posterior shifts by evidence weight. Self-explanation > a passive okay.
- 04
Misconception check
Conservative — only flags above 0.6 confidence. False positives erode trust.
- 05
Next-question generation
Pre-generates the next check, contextual to actual gaps.
§3 · Prerequisite graph
Concepts know what they need.
Auto-extracted. Teacher-overridable.
§4 · Intent-aware retrieval
Different questions.
Different strategies.
Definition, derivation, exercise — each routed differently.
§5 · The boundary
Stays inside
what you teach.
Enforced at the database, not the application.
§6 · Misconception detector
Wrong answers and wrong models.
Different things.
Conservative threshold. Teacher-trust comes first.
§7 · The operator layer
The school's controls. Exposed.
Cost rollup
Per-school, per-event-type, per-time-window.
Model dispatcher
Swap which model handles which task. No redeploy. Propagates in 60s.
Audit log
Every LLM call, embedding, retrieval — timestamped, costed, attributable.
Approval gates
Teacher onboarding, document upload, configuration changes.