Deterministic tasks
- JSON Schema validation
- Deterministic answer shuffling
- Game physics and canvas rendering
- Mastery and reward calculations
Case study · learning application and content quality
Level-Up Learning
I built Level-Up Learning for my son, with question-based missions, visible progress, and arcade rewards shaped by his feedback. As the original 565-question bank needed more variety, I built an offline AI generation pipeline, content validators, reusable React controls, and four game engines. The current build offers 3,158 playable missions, six-question practice sessions, and mastery based on independently solved questions, with no model calls during play.
Evidence and current limits
The September 7 source contains 3,249 question records, with 91 repeated prompts excluded from play while historical progress is retained. The resulting runtime serves 3,158 playable missions in 79 on-demand chunks, including 213 interactive questions. All 25 current Node tests passed during this review.
01 / Problem
Question variety had to grow without placing a model in the learning session.
A strong learner could quickly exhaust a small hand-authored question bank. Live inference would add latency, provider dependency, unpredictable output, and unnecessary data exposure during play.
The chosen architecture generates candidate content offline. Node.js validators treat every generated batch as untrusted input before approved questions and assets enter the deterministic browser application.
02 / Evolution
What changed as the project grew.
-
01
Working baseline
A React dashboard started with 565 missions, XP, and seven small games.
-
02
Agent content factory
Codex and Claude CLI lanes generated and reviewed structured question batches offline.
-
03
Admission gates
Schemas, duplicate checks, asset checks, and deterministic option shuffling controlled import.
-
04
Grown runtime
Chunked loading, interactive mission components, and local mastery state serve daily play.
03 / What I built
An offline AI content pipeline feeding a deterministic client application.
- Batch generation pipeline Offline CLI tools generate structured JSON mission batches with grade-level vocabulary constraints.
- Validator suite JSON schema validation, duplicate detection, and answer sanity checks reject invalid questions.
- Interactive game engines Tile Junction, Coordinate Quest, and 2 additional canvas/React engines power arcade rewards.
- Local progress state Mastery scores, earned XP, and streak tracking run client-side with zero external telemetry.
04 / Architecture
Offline AI generation feeds client-side deterministic runtime.
Offline Generation
Batch prompt scripts
Validation Gate
JSON Schema + duplicate check
Bundle Packaging
79 chunked JSON assets
Client Runtime
React + Arcade Canvas
User Mastery
Zero model calls in session
Level-Up Learning architecture separating offline AI authoring from the pure deterministic client game.
05 / System Boundary
AI writes candidate questions; code validates; browser runs the game.
Model reasoning
- Candidate question ideation
- Distractor option generation
- Explanatory hint drafting
Human decisions
- Curriculum topic selection
- Playtesting and son feedback review
- Visual theme approval
06 / Reliability
Complete isolation from external network dependencies during play.
The client application runs 100% offline once loaded. No API keys, no network timeouts, and zero risk of inappropriate model output in real time.
Batch import rejects any question with malformed JSON, duplicate question text, or invalid answer indices.
07 / Outcomes
3,158 verified missions with zero live model latency.
Content Scale
Expanded from 565 hand-crafted questions to 3,158 verified missions across math, science, and spatial reasoning.
Zero Playtime Model Cost
Eliminated ongoing API costs during daily use by moving all AI operations to build time.
Child-Safe Reliability
Deterministic content delivery guarantees no hallucinatory or inappropriate questions reach the screen.