Independent product case study / Education technology and adaptive practice
SparkTutor
SparkTutor is a privacy-conscious, rule-based adaptive learning prototype for mathematics and programming. It adjusts difficulty from learner performance, provides layered instructional support, and tracks chapter-level mastery in the browser without an account or a runtime generative-AI model.
01 / Brief
Learners need a clear next step, useful help after a miss, and progress that belongs to the topic they practiced. A trustworthy prototype also has to explain its adaptation, avoid overstating curriculum depth, and protect young users without adding account friction.
The experience uses large controls, friendly language, keyboard-visible focus, staged support, and no required profile. K–5 currently receives mathematics only; text-based programming begins in grades 6–8 and remains explicitly labeled beta.
02 / Architecture
Each stage has a bounded responsibility. The interface preserves source context and leaves the final judgment with the person using it.
Grade band selects a starting pace and the subjects available to the learner; K–5 is limited to mathematics in the current beta.
Grade → allowed subjects → default chapter → initial difficultyA deterministic TypeScript engine chooses a reviewed mathematics variation or a chapter-aligned programming checkpoint.
No browsing · no runtime model · no open-ended generated contentWhitespace, casing, numeric decimals, x = notation, and optional function parentheses are normalized before an accepted-answer comparison.
Transparent accepted formats · no symbolic algebra or code executionPerformance updates the current path only. Two correct answers raise the next level; a miss lowers it and opens supportive guidance.
Bounded 1–5 difficulty · path-scoped streak and masteryLearners move through a conceptual hint, first step, similar example, and full explanation after an attempt.
Reviewed strings tied to the current chapter · solution gatingA versioned learner record stores per-path totals in the browser, while reset controls support shared devices and privacy-aware recovery.
No account · no learner database · no runtime coach transmission03 / Product decisions
Tradeoffs made visible
The adaptive behavior is simple enough for a learner, caregiver, or recruiter to inspect. It does not claim predictive mastery or AI-driven pedagogy.
The coach stays inside local chapter content. This limits conversation variety but prevents invented explanations and keeps evaluation reproducible.
No sign-up reduces data collection and friction, but progress does not automatically travel between browsers or devices.
Rather than placing text-based variable and syntax questions in K–2, programming becomes available from grades 6–8. A future early-computing path would require visual, audio, and educator-reviewed interaction design.
Failure is a designed state
The learner state now hydrates from a versioned local key before persistence begins; browser tests verify that XP and path mastery remain after reload.
Records are keyed by grade, subject, and chapter. Switching paths shows separate accuracy, streak, level, and XP, and reset-this-path leaves other records intact.
The normalizer accepts documented numeric and syntax variants. Unmatched formats return a useful reason and the current conceptual hint instead of a bare wrong state.
Prompt-injection and off-topic categories return clear boundaries naming the current chapter. The coach cannot browse, reveal hidden instructions, or switch into an unrelated assistant.
04 / Evaluation
This is the evaluation surface documented by the implementation—not a claim that every target has already been met.
208 scripted interactions exercise four generated challenges across every currently advertised chapter, requiring complete prompts, hints, examples, explanations, accepted answers, and rejection of invalid input.
Twelve documented cases cover numeric decimals, equation notation, SQL functions with optional parentheses, capitalization, spacing, and representative programming formats.
Ten cases exercise prompt injection, answer extraction, off-topic health, news and finance prompts, hints, first steps, examples, and chapter-specific fallback.
Nine cases cover versioning, corrupt records, numeric bounds, path separation, invalid path removal, defaults, and enforced math-only K–5 state.
A release walkthrough verifies K–2 restrictions, correct grading, reload persistence, SQL normalization, mastery separation, hint gating, coach refusal, reset isolation, navigation, and public evidence pages.
ESLint, strict TypeScript, the automated suite, production build, custom-domain redirect, semantic structure, and desktop visual review must pass before deployment.
05 / Build ownership
AI accelerated drafting, implementation alternatives, and adversarial review. Product judgment, source selection, interface decisions, testing, and release approval remained mine.
Defined the adaptive rule: start by grade band, raise one level after two consecutive correct answers, lower one level after a miss, and remain within levels 1–5. Designed the learning-world navigation, challenge workspace, four-step hint ladder, path mastery board, and shared-device reset controls.
Exploring implementation options, generating draft test cases, checking edge cases, and accelerating repetitive code—always reviewed against the product boundary.
Chapter coverage, Answer equivalence. Failures were converted into visible recovery states and narrower claims.
Trust improves when missing evidence, partial provider failures, and the human decision are designed as first-class states instead of footnotes.
Choose an age-appropriate path → Generate a bounded challenge → Normalize and evaluate → Adjust one level → Reveal support progressively → Persist locally and visibly.
The learner state now hydrates from a versioned local key before persistence begins; browser tests verify that XP and path mastery remain after reload.
06 / Provenance
A primary reference for U.S. K–12 mathematics domains and progression terminology.
A primary reference for age-banded computer science concepts and practice boundaries.
Language behavior and foundational programming terminology.
Official Java language learning and platform references.
Primary SQL concepts including querying, joins, aggregations, and window behavior.
Web-platform terminology, semantic HTML, CSS, responsive design, and accessibility references.
Implementation note
Reviewed against 208 scripted chapter interactions, answer-equivalence and coach-boundary cases, versioned learner-state tests, the production build, and a critical browser workflow. This is product evaluation—not proof of educational outcomes, curriculum completeness, or accreditation.
See the system under real conditions