process_log.md

process_log.md — S2.T3.1 · AI Math Tutor for Early Learners

Candidate: John Eze
Challenge: S2.T3.1 — Tier 3
Date: 2026-04-24
Hard cap: 4 hours

---

Hour-by-Hour Timeline

Time	Activity
00:00–00:45	Read full brief end-to-end. Examined seed files (curriculum_seed.json, diagnostic_probes_seed.csv, parent_report_schema.json, child_utt_sample_seed.csv). Scaffolded repository structure: tutor/ package, data/, assets/, tts/, requirements.txt, README.md, SIGNED.md.
00:45–01:30	Built generate_curriculum.py (60+ item generator). Built tutor/curriculum_loader.py.
01:30–02:15	Implemented tutor/adaptive.py — BKT per sub-skill with update rules + Elo baseline. Built kt_eval.ipynb skeleton with held-out replay simulation.
02:15–02:45	Built tutor/visual.py (PIL counting images), tutor/feedback.py (EN/FR/KIN feedback table), tutor/asr_adapt.py (Whisper-tiny stub + langdetect).
02:45–03:15	Built tutor/db.py (encrypted SQLite store) and tutor/dp_sync.py (ε-DP aggregation design).
03:15–03:45	Built demo.py (Gradio child-facing UI: learner select, tutor loop, parent view).
03:45–04:00	Built parent_report.py. Completed footprint_report.md, README.md, process_log.md. Final push.

---

LLM / Tool Use Declaration

Tool	Why used
GitHub Copilot (Claude Sonnet 4.6)	Code scaffolding, boilerplate generation, docstrings, README drafting. All logic reviewed, debugged, and adapted by me.

Three sample prompts I actually sent

1. "Build the BKT update rule for 5 sub-skills — P_know, P_learn, P_guess, P_slip — with a mastery threshold of 0.85 and return the next best item."
2. "Write a PIL function that renders N emoji-like circles on a white canvas and saves to assets/, returning the image path and the correct count."
3. "Generate a Gradio demo with three tabs: learner avatar select, tutor question-answer loop with image display and mic/text input, and a password-protected parent report view."

One prompt I discarded and why

I drafted a prompt asking Copilot to "fully fine-tune TinyLlama with QLoRA for numeracy feedback generation". I discarded this because (a) 4-hour time cap, (b) TinyLlama int4 GGUF is ~650 MB which blows the footprint budget if bundled, and (c) rule-based + template feedback is faster, more predictable, and fully explainable for a child-safety product. LLM enhancement can be added post-hackathon.

---

Single Hardest Decision

The hardest decision was choosing between a Deep Knowledge Tracing (DKT) GRU model and Bayesian Knowledge Tracing (BKT). DKT would likely score higher AUC on complex interaction sequences but requires PyTorch, adds ~50 MB to the footprint, and takes 30+ min to train even on CPU. BKT is interpretable, has zero training cost, runs in microseconds per update, and is well-understood in educational research. I chose BKT as the primary model and kept the Elo baseline for comparison in kt_eval.ipynb — this satisfies the brief's AUC comparison requirement while staying within the 75 MB footprint and 4-hour time budget.

---

Log updated continuously throughout the session.