Artstache: Why a Graph Beats a Search When the Order Matters
Parents don’t default to “that’s nice!” because they don’t care — they do it because after a 10-hour workday they are exhausted, holding a curling drawing at the kitchen counter. And they don’t know what else to say. Children’s drawing develops in stages, in order. A normal AI search can’t tell a 4-year-old’s tadpole person from a 10-year-old’s comic — the descriptions sound the same. Artstache stores Viktor Lowenfeld’s developmental stages (1947) as a chain in a database (Neo4j), so the order can’t be skipped, and gives the parent a short guide — the Ghost Guide — they can read in about a minute.
The Kitchen Counter Problem: Guilt, Clutter & Fatigue at 6:15 PM
Founder's Truth: I'm not a parent. I built Artstache for my sisters and friends — working parents who love their kids' art but after a 10-hour day default to “that's nice!” and feel guilty tossing the paper pile at midnight. I earned my design degree much later than most, and learning visual literacy helped me cut through noise. This is my contribution.
Child development literature is dense and academic. Parents want to guide creative growth but have zero time for lesson prep. Result: 25+ unsorted drawings piling up on counters, guilt-driven midnight purges, and passive iPad screen time.
“A tired parent gets something meaningful and stage-accurate to say, in about a minute. Web-first, one-handed scan at the kitchen table. Out of scope: classroom dashboards, selling art supplies.”
Why “What Is It?” Backfires: The Tadpole Person
I went through public school art curricula. Most of them are already built on Viktor Lowenfeld’s 1947 developmental stages. The research was finished decades ago — what was missing was a way to use it in about a minute, at the kitchen table:
“A tadpole person is not a mistake — it is how a young child’s mind works. Artstache’s job is to translate the drawing, so a parent can celebrate what is emerging instead of correcting the anatomy.”
Why a Graph Beats a Search When Order Matters
A normal AI search matches by similar words. A 4-year-old’s tadpole person and an older child’s stick-figure comic get described with the same words — “figure, limbs, face” — so the search mixes up the stages and can recommend a perspective lesson years before a child is ready for it.
In child development, order is everything. You can’t teach perspective before a child is drawing a baseline — the stages have to come in sequence.
If the lookup finds no matching stage, or the AI’s confidence is under 0.75, the app stops and shows no milestone badge. The AI can only write from the exact stages the lookup returned — it can’t invent a stage.
The No-Line Rule: From Clinical to Archival
What’s Proven, and What Isn’t
My rule: No fake user counts. I only report what I can check in the code and the tests. The biggest risk in this product is an AI suggesting a stage out of order — so that is what the tests hammer.
milestone_guard.ts)✦See the Guardrail Code (milestone_guard.ts)▼
// The research sets the order; the AI only writes the words
const result = await neo4jDriver.executeRead(async (tx) => {
return tx.run(`
MATCH (current:Stage {id: $stageId})-[:PRECEDES]->(next:Stage)
RETURN current, next
`, { stageId });
});
// If the check fails, show nothing — no badge is better than a fake badge
if (!result.records.length || confidence < 0.75) return null;
// The 5-step workflow only writes from stages that checked out
const guide = await langGraphDAG.invoke({
stage: result.records[0].get('current'),
next: result.records[0].get('next'),
drawingPrimitives: visionFeatures
});latency_runs.jsonl). A full scan: ~800ms typical. The free database sleeps after 72 hours idle, so the first scan after that takes ~30 seconds unless the keep-alive is running. No production monitoring yet.- Graphs beat searches when order matters: A search matches a 4-year-old’s tadpole person to a 10-year-old’s comic because the words are similar. Development is a sequence — storing the stages as a chain enforces the real order, not a similar one.
- I over-built the agent team: 11 named agents felt clever, but 2 of the steps do most of the work. If I’d had a PM, they would have cut the rest on day 2. Don’t name a step unless you can justify it to a tired parent.
- Shame is the real competitor: Not other ed-tech apps. A parent feeling judged by a clinical UI and defaulting to “that's nice!”. The No-Line Rule was a direct response to shame.
- What I'd do differently: Move the vision step onto the phone itself (Wasm/ONNX), so scans work in basements and cabins with no signal.
- Pilot status, honestly: The app is live and the stage-order checks are tested. I’m recruiting 10–15 pilot families (kids ages 3–8) for real kitchen-table trials — 0 enrolled yet, so no claims about results for kids.
- Still writing: The short follow-up activities — 15-minute activities across all 6 stages, about 15–20 milestones in total.
Shipped, Verified, and Honest About What's Next
I built Artstache for real kitchen tables. There are no user counts to report yet — no families have enrolled. What I can report: the stages live in a graph that keeps them in order, the tests prove the order holds even when I try to break it, and the design feels like a family archive instead of a medical chart.