Don —
Live • artstache.app →
01 Discovery/02 Research/03 Systems/04 Craft/05 AI Leverage/06 Outcome
GRAPH RAG • MILESTONE BADGE • ZERO-PREP • SCIENCE-GROUNDED • COPPA • DESKTOP-FIRST

Artstache: From Kitchen Clutter to Archival Monograph

Parents don’t say the wrong thing because they don’t care — they say it because they don’t know what else to say. Artstache models Lowenfeld's developmental stages (1947) alongside CDC and NAEYC milestones into a property graph, generating a 60-second Ghost Guide at the kitchen table grounded in deterministic relationships so it never hallucinates milestone progression.

is_grounded_in_neo4j gateNo-Line Rule • 4 paper elevations48px thumb targets • 1-tap guide
View Live at artstache.app →
Every mark is a milestone
Project Meta • Gen-Pop
RoleMulti-disciplinary: Research → Product → Design System → Code
ScopeDiscovery → Shipped • 0-to-1
StackNext.js 16, Neo4j AuraDB, LangGraph, Gemini 2.5 Flash, Tailwind JIT
Privacy80KB Canvas buffer, EXIF stripped, zero facial storage
Outcome<800ms target SLA, [:PRECEDES] graph order
Why This Matters to Any Recruiter / Hiring Manager
This shows I can take validated domain milestones, model them as a property graph, and ship them as a zero-prep consumer product parents can use in 15 minutes. Same method works for any complex domain — media metadata, brand archives, enterprise knowledge graphs, or workflow tools where trust matters more than speed.
01 / DISCOVERY

Research Drain + Missing the Click

Two problems I saw at kitchen tables, not in analytics.

Problem A • Research Drain
Googling child psychology at 6:15 PM doesn't help.
Parents are in the 15-minute dinner window, kid hands them a drawing, they want to say something meaningful but default to "What is it?" — well-meaning but conversation-ending. No time for research.
Artifact: 6:15 PM journey map
Problem B • Missing the Click
"What is it?" shuts down the moment.
Every mark is a developmental milestone, but parents don't have the language. We need a Ghost Guide — 1-tap, science-backed, reassuring, not clinical. Like a museum label for their kid's art.
Artifact: Ghost Guide concept
02 / RESEARCH

Lowenfeld as Ontology, Not Moodboard

From 80 years of developmental psychology to deterministic graph triples.

Domain Modeling • Lowenfeld 6 Stages → Graph Triples
What I Did
Synthesized Lowenfeld, CDC, NAEYC into 6 stages → [:PRECEDES] chain. Each node: (art_trait, milestone, citation, confidence). No vectors, only triples. Graph enforces no skipping — 3-year-old can't be in Schematic without Preschematic.
Lowenfeld OntologyCDC • NAEYC CorpusCOPPA Review
Graph Schema
(:ArtTrait)-[:INDICATES]->(:Milestone)
(:Milestone)-[:PRECEDES]->(:Milestone)
(:Milestone)-[:CITED_BY]->(:Source)
No badge > fake badge. Confidence < 0.75 = null, not guess.
03 / SYSTEMS

GraphRAG > VectorRAG for Milestones

5-Node Village DAG • LangGraph
01
Extract
02
Ground in Neo4j
03
Gate
04
Wrap Tone
05
Guide
Deterministic Guardrail — Never Hallucinate a Milestone
const result = graphRag.query({
  embed: artEmbed,
  corpus: childDevCorpus, // CDC, NAEYC only
  mustCite: true
});

if (result.confidence < 0.75) return null; // No badge > fake badge

return llm.wrap({
  grounded: result.milestone,
  tone: persona === "parent" ? "reassuring" : "standards"
});
Math does math, LLMs do context. Graph for truth, LLM only for tone.
04 / CRAFT

Museum Archive, Not Medical Chart

Design System • No-Line Rule + 4 Paper Elevations
No-Line Rule: No clinical borders, only paper elevations (shadow + bg) — feels like archival monograph, not SaaS.
Dual-voice typography: Instrument Serif for narrative, JetBrains Mono for meta, Newsreader for body — parent vs standards.
48px thumb targets, 1-tap Ghost Guide: Designed for one-handed use while holding kid art in other hand. Zero-prep.
Token Specimen
#FAF9F6 — paper
#141413 — ink
#FF4D23 — signal only
AA 4.5:1 minimum • Mono 12/14 for tables • Inter 13.5 for UI
05 / AI LEVERAGE

Privacy First, Then Speed

Privacy Architecture
80KB in-browser Canvas buffer, EXIF stripped in-browser, zero facial storage, Canvas memory only — designed with COPPA principles in mind, not certified COPPA compliant. No biometric PII stored.
Artifact: In-browser Canvas memory
Latency Profile
Sub-40ms Cypher query verified locally via telemetry. End-to-end target <800ms SLA, dependent on Gemini Flash + Village DAG. AuraDB free-tier sleeps after 72h (cold start adds ~30s unless watchdog active). Measured via local logs.
Artifact: Sub-40ms Cypher verified locally
06 / OUTCOME

Shipped, Verified, Honest About What's Next.

No vanity metrics yet — early stage, honest pilot. The outcome is that a complex developmental ontology now runs as a deterministic graph with a tactile design system that feels like a museum archive, not a medical chart.

Design Outcome
No-Line Rule + 4 paper elevations + dual-voice typography + 48px thumb targets + Ghost Guide. From Research Drain to Kitchen Table Studio.
Technical Outcome
Neo4j AuraDB with Lowenfeld (1947) triples, GraphRAG structural ordering, 5-node Village DAG, sub-40ms local Cypher, and in-browser Canvas privacy architecture.
What I'd Measure Next
Time to guidance, trust in milestone badge, reduction in tab switching, qualitative feedback on whether Ghost Guide replaces "What is it?" with meaningful dinner conversation. Recruiting 10-15 families ages 3-8 for real kitchen trials.
Outcome — Verified in Code, Not Vanity Metrics
Live • Honest Pilot • No Synthetic Metrics
<800ms
Target SLA
Sub-40ms Cypher verified locally via telemetry. E2E depends on model load.
[:PRECEDES]
Structural Order
Graph structurally prevents out-of-order milestones. Null returned if confidence < 0.75.
80KB
Privacy buffer
In-browser Canvas memory • EXIF stripped • Zero biometric PII stored.
What I Verified in Code
  • Determinism: Graph enforces [:PRECEDES] relationship — structurally prevents out-of-order milestones. is_grounded_in_neo4j gate returns null when confidence <0.75, rather than synthetic badge. Verified via golden set of 50 pedagogy prompts (local), not production user data.
  • Privacy Architecture: 80KB in-browser Canvas buffer, EXIF stripped in-browser, zero facial storage, Canvas memory only — designed with COPPA principles in mind, not certified COPPA compliant. No biometric PII stored.
  • Latency Profile: Sub-40ms Cypher query verified locally via telemetry. End-to-end target <800ms SLA, dependent on Gemini Flash + Village DAG. AuraDB free-tier sleeps after 72h — cold start adds ~30s unless watchdog pinging. Measured via local logs, not production APM.
Next Steps — Generalist Roadmap
  • Content Synthesis: Reduced manual mapping of traits → milestones from ~3hrs per stage (manual research) to ~15min assisted via LLM APIs + prompt design — measured during authoring of 6 stages, not production workload.
  • Authoring 15-min micro-sprint curriculum across 6 Lowenfeld stages (~15-20 milestones total).
  • Wasm/ONNX edge vision — 120ms offline, zero-cloud for rural WiFi.
  • Recruiting 10-15 families ages 3-8 for real kitchen trials — 0 enrolled yet, honest.
Transferable pattern: Asset metadata as property graph, creator workflows as time-budget workspaces, deterministic grounding where hallucinations would harm trust. Same architecture works for media archives, brand asset management, or any enterprise knowledge graph.
Let's Talk • Available in Duarte, CA

Building something where trust matters more than speed?

I'm a multi-disciplinary designer — Product, Brand, Design Systems, Code. I partner across the stack from discovery to live production. Based in Duarte, CA (91010) — available for hybrid in Glendale, remote, or on-site. 0-2 weeks notice.

Email me →LinkedIn →GitHub →
What I bring
• Product, Brand, Design Systems, Code
• Figma → Next.js 16 → Prod
• GraphRAG, LangGraph, Neo4j
• 0-to-1 to scale, 10+ years
don-the-imaginator.com • #FAF9F6 / #141413 / #FF4D23 • 0.00 CLSBuilt with Next.js 16, Tailwind JIT
Next Case Study • Configurable
Desktop Enterprise • Ships Daily4.2s scan • $18/mo

3PM Daily: Closing Auction Debrief Terminal

How I turned 800+ tickers into 4 setups in 5 minutes — desktop-first, 5-archetype tokens, TA-Lib guardrails.

View 3PM Daily →All Work →
© Don Martirez — Artstache Gen-Pop • Multi-disciplinary • #FAF9F6 / #141413 / #FF4D23 • 0.00 CLS • COPPA