itsazza/KTT-Day3

---

pipeline_tag: automatic-speech-recognition tags: - offline - edge-device - gradio - education

🧠 Edge-Optimized Math Tutor Orchestrator

Model Description

This repository contains the orchestration and adaptive logic layer for an offline, CPU-bound AI Math Tutor. Rather than acting as a standalone monolithic weight file, this pipeline links extreme-edge open-source models with Bayesian logic to meet a strict < 75MB operating footprint constraint.

Core Architecture

1. ASR Component: openai/whisper-tiny (39M parameters). Handles primary transcription and number extraction.
2. Adaptive Brain: A bespoke Bayesian Knowledge Tracing (BKT) engine. Evaluates real-time P(mastery) against 5 distinct mathematical sub-skills to dynamically filter question difficulty.
3. Visual Grounding: Procedural geometry generation (Pillow) entirely replacing traditional object-detection (e.g., OwlViT) to maintain the disk footprint and guarantee < 2.5s inference latency.

Code-Switching Capability

The orchestrator includes custom logic layered over the ASR to detect and gracefully handle language code-switching (e.g., a child being prompted in Kinyarwanda but answering in English).

Evaluation

Tested against an Elo-style baseline on a held-out replay of 100 historical simulated student interactions.

• BKT Model AUC: Consistently outperforms standard Elo algorithms in predicting next-response correctness due to faster adaptation to early-learner "slip" and "guess" parameters. (See kt_eval.py in the linked GitHub repo for exact metrics).

Limitations

Due to the strict 75 MB footprint constraint, this pipeline does not utilize quantized LLM adapters (QLoRA) for language generation, relying instead on high-quality, pre-generated localized stems and rule-based regex extraction.