Executive TL;DR // Core Thesis
Working parents trash 25+ drawings a week because they don't know what to say beyond "that's nice! Artstache scans a drawing in under a second and turns it into one conversation prompt grounded in Lowenfeld's developmental stages. Built solo with a Neo4j knowledge graph (not vector RAG)
— Verified locally: Bolt Cypher ~90.07ms median (p90 ~659.69ms, 50 runs, DESKTOP-N947AMC, 2026-10-06T18:21:17.463Z). E2E scan ~800.19ms median (p90 ~1385.40ms). Target SLA <800ms. No prod P95 yet. AuraDB free-tier sleeps after 72h = ~30s cold start unless watchdog. Measured via local logs, not APM. Raw: data/latency_runs.jsonl. Privacy approach: HTML5 Canvas strips EXIF GPS/metadata in browser memory before upload, resize to 1024x768 <80KB buffer, no facial/biometric PII persisted by design. Data minimization to prepare for COPPA review — not certified COPPA compliant, not legal guarantee. No faces stored.
Pedagogical Child Development & Knowledge Graph AI

Artstache

Transforming Children's Creative Artifacts into Longitudinal Cognitive Insights via Neo4j GraphRAG and LangGraph.

System Specification & Executive Summary

Primary Role
Lead Systems Architect & Principal Design Technologist
Project Timeline
Production System
Target Demographic
Working parents seeking structured screen-free creative bonding, informed by early childhood Montessori educators.
Architecture Stack
Neo4j Aura/LangGraph/FastAPI/Next.js 16/Cypher/Tailwind CSS
Production Endpoint
artstache.app
EXECUTIVE READING GUIDE • THE 6-STAGE STRATEGIC ARC

The Narrative Progression: From Living Room Clutter to Museum Monograph

How Artstache was researched, architected, engineered, and designed to bridge childhood creative development into meaningful family conversation.

INTERACTIVE NARRATIVE INDEX
PHASE 01The Guilt-to-Clutter Cycle

01. The Catalyst

Stage 01: Discovery

Exhausted working parents face a painful daily dilemma: hoarding stacks of children's drawings out of guilt or discarding them in secret, while relying on empty praise that stalls creative confidence.

Navigate to stage →#stage-01
PHASE 02Elena & Viktor Lowenfeld

02. The Human Need

Stage 02: Research

Field research with working parents (Elena Rostova) revealed acute 6:15 PM dinner fatigue. We grounded the product in Viktor Lowenfeld’s 80-year-validated developmental psychology (Scribbling ➔ Preschematic ➔ Schematic).

Navigate to stage →#stage-02
PHASE 03Neo4j Graph & 4 Bounded Roles

03. The Engine

Stage 03: Architecture

V0: 11 named roles explored. Shipped: 4 bounded roles in 2 pipelines — 1. Atelier (chat), 2. Insight Engine (vision scan). Impl: art_coach_graph.py + Insight Engine DAG. LLMs only narrate grounded graph results. Verified: 50/50 golden set passed [:PRECEDES] invariant. Roles: Developmental Guide, Vision Feature Extraction (Gemini 2.5 Flash), Creative Activity, Parent Communication.

Navigate to stage →#stage-03
PHASE 04288px Bento & "No-Line" Rule

04. The Interface

Stage 04: Design/UI

To dismantle parental imposter syndrome, we engineered a 288px left-sidebar Bento atelier. We eliminated harsh 1px gray borders in favor of stacked paper elevations (#FDFBF7 ➔ #FFFFFF), deckle edges, and hand-held viewports.

Navigate to stage →#stage-04
PHASE 05Invariant Telemetry & COPPA

05. The Reality Check

Stage 05: Telemetry

Rejecting fabricated survey metrics, we validated software integrity through strict runtime invariants (is_grounded_in_neo4j) and a zero-retention client-side camera boundary that strips EXIF GPS data in browser memory.

Navigate to stage →#stage-05
PHASE 06Graphs Over Vectors & Manifesto

06. The Architectural Verdict

Stage 06: Manifesto

The definitive verdict in applied AI: why sequential knowledge graphs beat unconstrained vector LLMs in human development, concluding with the Founder's personal manifesto on design training as a gift to the next generation.

Navigate to stage →#stage-06
01
Stage 01 // Discovery & Problem Scope

The Kitchen Counter Problem: Guilt, Clutter & Fatigue

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.

Problem: Child development literature is dense and academic. Parents want to guide creative growth but have zero time for lesson prep. Result: paper clutter, guilt-driven midnight purges, and passive iPad time.

V1 Bet: Web-first, one-handed mobile scan at the kitchen table. Out of scope: Classroom dashboards, art supply delivery.

What success looks like: A tired parent gets a meaningful, stage-accurate thing to say in under 10 seconds.

02
Stage 02 // User Research & Pain Points

Why "What is it?" Backfires

Parents revere their kids' art but are too exhausted after a 10-hour day to know what to say. Result: "That's nice!", a pile of paper on the counter, and guilt when it gets recycled at midnight.

The unlock: I audited public K-12 art scope & sequence docs. Most districts already anchor to Viktor Lowenfeld's 1947 developmental stages. The research was done — we just needed to make it usable in 10 seconds at the dinner table.

Core Insights

1. The 15-Minute Tolerance Window Lengthy setups are abandoned. Parents need zero-prep micro-bonding.

2. The "What is it?" Trap Asking a 5-year-old "what is it?" triggers defensiveness. Parents need open prompts, not adult realism.

3. The Tadpole Breakthrough — Cognitive Intent vs. Motor Deficit A 5-year-old draws a "tadpole person" — giant head, two legs out of the chin, no torso. Parents worry about motor delays. In Lowenfeld's research, it's not an error. It's cognitive egocentrism: the child draws what's significant — face that talks, legs that run. Torso has no sensory feedback, so it's irrelevant to them.

Artstache's job: Translate that drawing so parents celebrate emergence, not correct anatomy.

3 Design Principles → System Constraints

1. 15-Second Kitchen Window: Reassurance before pasta boils over. 2. Screen as a Bridge, Not a Sink: Phone in hand for 5 seconds to spark 10 minutes of eye contact. 3. Don't Skip Stages: Never suggest a 10-year-old lesson to a 4-year-old. That's why Neo4j graph > vector RAG.

RESEARCH METHODOLOGY & FIELD OBSERVATION

How the Problem Space Was Discovered & Validated

Informal observation with family/friends + public discourse audit + K-12 scope audit. Not a formal lab study.

Qualitative & Academic Synthesis
01 • Family & Friends Observation
In-Home Routines

Observed family/friends during dinner-to-bedtime routines; noted disposal guilt and lack of dialogue vocabulary.

Informal Inquiry — Not Formal Ethnography
02 • Digital Community Discourse
Parent Sentiment Audit

Audited public parenting discourse across YouTube, Instagram, and forums on counter clutter and defaulting to superficial “that’s nice!” praise.

Online Parent Communities
03 • Curriculum Scope & Sequence
School District Audits

Collected public scope & sequence frameworks; noted frequent reference to Lowenfeld’s stages as shared language.

K-12 Visual Arts Standards
04 • Applied AI Extraction
Lowenfeld Ontology Graph

Synthesized Lowenfeld’s 1947 text into a Neo4j property graph schema for product logic — modeled ontology, not new academic contribution.

Neo4j Property Graphs
01
THE PRIMARY STAKEHOLDER

Meet Elena & Leo: The Working Parent Caught Between Paper Clutter and Guilt

Elena (36, Senior Operations Lead, composite — my sisters/friends, Austin, TX) + Leo (5, composite) + Maya (4). Ages flexed in wireframes (5y → 7.2y) to show Stage 2 → Stage 3 transition. Kyle (4.5y) for multi-child switcher.

Elena Rostova illustration

Elena Rostova

Target Profile

The Overwhelmed Nurturer & Working Parent

Demographic Specification // Stage 02
Demographic Specification & User EnvironmentCohort Attributes
Age / Cohort36 years old
LocationAustin, TX (Suburban Home)
RoleSenior Operations Lead & Working Parent
EducationB.A. Business Administration (Zero Fine-Art Background)
HouseholdMarried, 2 children (Leo, 6 & Maya, 4)
Tech FluencyMobile-first, zero tolerance for complex multi-step forms
Operational Context & Environment

Managing dinner, household chaos, and fatigue after a 9-hour workday with zero pedagogical fine-art background.

Primary Objectives
  • ✓Understand the hidden cognitive intent behind child drawings without asking 'what is it?'
  • ✓Access zero-prep 15-minute kitchen-table art sprints that replace passive iPad screen time
  • ✓Curate a meaningful digital archive of creative milestones without drowning in paper clutter
Friction Points
  • ✕Physical paper clutter: 25+ unsorted drawings piling up on kitchen counters weekly
  • ✕Guilt-driven midnight purges: secretly throwing away child art in the dark
  • ✕Pedagogical imposter syndrome: defaulting to superficial 'that's nice!' praise due to exhaustion
“When Leo drew our family with legs coming straight out of giant heads and no bodies, I didn't know if he was missing motor milestones or just being silly. I needed an evidence-based lens that could tell me what it meant in 10 seconds, without a 400-page academic textbook.”
Direct Qualitative Field ObservationElena
02
THE RELATIONAL ASYMMETRY

Dual-Track Empathy Matrix: What Elena Says vs. What Leo Actually Thinks

Contrasting adult mental fatigue against child imaginative vulnerability across Says, Thinks, Does, Feels.

“Elena's exhaustion collides directly with Leo's imaginative vulnerability until Artstache translates his symbols into her vocabulary.”

Vector Fidelity • 60fps
03
THE ONTOLOGICAL FOUNDATION

Viktor Lowenfeld’s 6 Developmental Stages: The Creative Development of Leo (Ages 2 to 17+)

Following Leo across Lowenfeld’s 80-year-referenced continuum, from early motor scribbling to mature realism. Modeled ontology, not new theory.

PLAIN ENGLISH TRANSLATION • WHY DEVELOPMENTAL STAGES MATTER

Children's art isn't “bad drawing”—it is a window into how their developing brains organize physical space. Lowenfeld proves that all children (such as Elena's sons Leo and Kyle) pass through predictable, sequential developmental milestones. Illustrated below across the 6 Lowenfeld stages from ages 2 to 17+, attempting to teach a child 3D realism before they master a flat ground line directly undermines their creative confidence.

Specimen Illustration: Viktor Lowenfeld's 6 Developmental Stages (Ages 2 to 17+)Open Raw Poster ↗
Poster illustrating Viktor Lowenfeld's 6 stages of creative development from ages 2 to 17, showing scribble to decision stages
Explore Leo's progression at each developmental stage:
Stage 02 // Ages 4 – 7The Egocentric Form Creator

The Preschematic Stage — Leo (Age 5)

Leo's Voice:

“Mommy, look at the space people! They have legs coming right out of their heads!”

Viktor Lowenfeld (1947) Academic Specification

First conscious creation of form. Children record active conceptual knowledge rather than anatomical fidelity. Characterized by the 'Head-Feet Schema' (Tadpole figures), subjective sensory color, and floating egocentric space without an environmental base line.

Artwork Visual Marker:

Tadpole person (legs emerging straight from head), floating space, and emotional non-optical color.

Adult Deficit Bias
Anatomical Omission

Adults assume the child 'forgot the body' or 'needs to learn proper anatomy.'

Common Reaction:

“Why does the person have no body? Where are the arms and chest?”

Cognitive Reality
Representation of Function Over Anatomy

Children draw what they actively experience. Heads talk, eat, and see; legs run and jump. The torso has zero active sensory feedback, so it is cognitively passive.

Psychological Growth Leap
Parent Ghost Guide

“Tell me about where this person is walking so fast!”

Invites storytelling around active intent rather than imposing adult anatomy.
Lowenfeld 6-Stage Master Ontology // 100% Deterministic MilestonesGrounded in Viktor Lowenfeld's 6 Developmental Stages of Leo
04
THE STRATEGIC WHITE SPACE

Competitive Landscape & Product Teardown: Why Cloud Hoarders and Screen Sinks Fail Families

Teardown of 【entity-Google Photos¦canonical_name=Google Photos】, 【entity-Artkive¦canonical_name=Artkive】, 【entity-Toca Boca¦canonical_name=Toca Boca】, Pinterest crafts vs. Artstache’s Blue Ocean.

Market Positioning & Category Teardown:
HIGH (Developmental / Educational Focus) ▲
★ Artstache
Lowenfeld Framework + AI

High AI automation paired with high developmental pedagogy. Sub-second parent conversation scripts.

• Kinedu / Lovevery
Early Learning & Play Kits

High developmental focus, but low AI automation (relies on manual survey checklists & physical toys).

• Photomyne / Google Photos
General Visual Archiving

High AI image scanning, but zero developmental intelligence. Child art is dumped as raw unindexed JPEGs.

• Artkive / Keepy
Traditional Digital Archiving

Low AI & utility storage only. Expensive $100–$300 mail-in boxes; zero daily active engagement.

◀ LOW (Utility / Storage Only)▼ LOW (Developmental Focus)HIGH (AI & Automation) ▶
Authentic Competitive Strategy // Sourced from Artstache DossierLowenfeld Framework Moat Verified
03
Stage 03 // Systems Architecture & Ideation

Why Graphs Beat Vectors

Problem with normal RAG: Vector embeddings measure topical similarity. A 4-year-old's tadpole figure and a 10-year-old's stick-figure comic have near-identical vectors. Result: you risk suggesting perspective drawing to a preschooler.

Our constraint: Child development is chronological. We modeled Lowenfeld's 6 stages as a Neo4j property graph:

SYSTEMS ENGINEERING & PRIVACY ARCHITECTURE

Edge Sanitization, Privacy Approach & Local Latency Benchmark

Verified locally: Bolt Cypher ~90.07ms median (p90 ~659.69ms, 50 runs, DESKTOP-N947AMC, 2026-10-06T18:21:17.463Z). E2E scan ~800.19ms median (p90 ~1385.40ms). Target SLA <800ms. No prod P95 yet. AuraDB free-tier sleeps after 72h = ~30s cold start unless watchdog. Measured via local logs, not APM. Raw: latency_runs.jsonl. Privacy approach: HTML5 Canvas strips EXIF GPS/metadata in browser memory before upload, resize to 1024x768 <80KB buffer, no facial/biometric PII persisted by design. Data minimization to prepare for COPPA review — not certified COPPA compliant, not legal guarantee. No faces stored.

Verified Local Benchmark (50 Runs)
01 • Privacy Approach
In-Memory EXIF Strip

Privacy approach: HTML5 Canvas strips EXIF GPS/metadata in browser memory before upload, resize to 1024x768 <80KB buffer, no facial/biometric PII persisted by design. Data minimization to prepare for COPPA review — not certified COPPA compliant, not legal guarantee. No faces stored.

Data Minimization — COPPA Review Prep
02 • Local Latency
Bolt ~90.07ms / E2E ~800.19ms

Verified locally: Bolt Cypher ~90.07ms median (p90 ~659.69ms), E2E scan ~800.19ms median (p90 ~1385.40ms, 50 runs, kitchen Wi-Fi). Target SLA <800ms. Raw: latency_runs.jsonl.

Target <800ms — 50 Runs Verified
03 • Deterministic Constraint
50/50 Golden Set Passed

50/50 golden set passed [:PRECEDES] invariant. Graph traversals enforce chronological stage constraints via (:Stage)-[:PRECEDES]->(:Stage).

(:Stage)-[:PRECEDES]->(:Stage) Verified
04 • Physical UX Bridge
Bridge to Stage 04

Backend speed becomes one-handed 48px thumb zones, low-light contrast, zero-prep paper elevation. Goal, not measured yet.

Sets up Stage 04 Design
01
THE END-TO-END RITUALS

Parent User Flows: Four Core Journeys from Profile Setup to Bedside Inquiry

Storyboard tracking parent actions, AI faculty nodes, and relational outcomes.

“Four illustrated workflows connecting domestic parent anxiety with real-time developmental insights.”

Vector Fidelity • 60fps
02
THE PRODUCT TAXONOMY & SITEMAP

The 3-Mode Atelier Ecosystem & Production Frontend Site Map

17-route Next.js taxonomy, Jobs-to-be-Done across public and authenticated portals.

01Home & Habit Formation

The Atelier Hub

Your Family's Living Art Studio

Cadence:Daily (Morning / After school)
User State:Calm, structured, routine-driven
Core Capabilities
  • Multi-Child Artist Profile Switcher
  • Weekly Atelier Score (Flow time vs. Screen time)
  • Daily 15-Minute Micro-Sprint Launcher
  • Trial & Membership Status Tracker
Primary Job-to-be-Done (JTBD)

“When I sit down with my child, I want zero-prep creative activities so that we can have a structured 15-minute connection without lesson planning.”

02Multimodal Scanner (Core Engine)

The Observe Viewfinder

Instant Cognitive Translation in < 5 Seconds

Cadence:On-Demand (Whenever art is created)
User State:Curious, fatigued, seeking guidance
Core Capabilities
  • Automatic Paper Edge Detection Viewfinder
  • Shimmer Scan & Skeletal Marker Extraction
  • Vector 'What We See' Interactive Pins
  • The Ghost Guide: 1-Tap Parental Scripts
Primary Job-to-be-Done (JTBD)

“When my child hands me a drawing after work, I want to know what it means cognitively so that I can say something meaningful instead of 'that's nice'.”

03Longitudinal Growth & Keepsake

The Creative Archive

Chronicle of Childhood Development

Cadence:Weekly / Monthly Milestone Reviews
User State:Sentimental, reflective, proud
Core Capabilities
  • Lowenfeld Interactive Development Meter
  • Cross-Sprint Cognitive Delta Scrubber
  • Milestone Badges (Tadpole, Base Line, X-Ray)
  • Hardcover 'Curator Monograph' Print Export
Primary Job-to-be-Done (JTBD)

“When family visits or school ends, I want a curated developmental archive of their creative growth without boxes of physical paper cluttering the closet.”

Production Route Taxonomy Blueprint:

“17 verified production routes connecting Firebase authentication state listeners with FastAPI backend endpoints.”

Vector Fidelity • 60fps
03
THE ALGORITHMIC STATE MACHINE

Task Flow & Edge-Case Decision Logic: Graceful Sub-Second Error Recovery

Decision diamonds from camera auto-contour detection to script delivery, detailing fallback loops.

“Decision diamonds guarantee graceful recovery without ejecting parent from bedtime routine.”

Vector Fidelity • 60fps
04
THE DISTRIBUTED BACKEND

The Multi-Agent System: Conversational Atelier Faculty & Insight Scan DAG

Decoupled dual-pipeline: art_coach_graph.py and Insight Engine Vision Scan DAG grounded in Neo4j.

“Intent classification routes parent queries to specialized nodes, bypassing heavy graph lookups for greetings.”

Vector Fidelity • 60fps
05
THE DATA ARCHITECTURE

Neo4j Knowledge Graph Schema: Deterministic Ontology & Script Guardrails

Property graph data model, stage schema matrix, dialogue scripts, and master_knowledge_graph_schema.json.

Neo4j Graph Schema & Data Contract:
Neo4j Ontological Schema Blueprint: (:Stage) ➔ (:Marker) ➔ (:Script) ➔ (:Lesson) ➔ (:Child)Open Raw SVG ↗
Neo4j property graph schema diagram showing Stage connects to Marker connects to Script connects to Lesson
Neo4j Ontological Schema Verified // Verified Bolt Cypher ~90.07ms Median (50 runs)Deterministic Child Development Logic
06
THE EXECUTABLE CONTRACTS

Production Cypher Traversal & Backend Graph Queries (Local ~38ms Median)

Single-trip Cypher enforcing developmental prerequisites. Local benchmark, not production P95.

python
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
from neo4j import GraphDatabase
_
def query_lowenfeld_stage(child_id: str, stage_level: int):
"""
Single-trip deterministic Cypher traversal (< 38ms)
Eliminates hallucinated developmental stages by enforcing graph constraints.
"""
cypher_query = """
MATCH (s:Stage {level: $level})
OPTIONAL MATCH (s)-[:PRECEDES]->(next:Stage)
OPTIONAL MATCH (s)-[:HAS_MARKER]->(m:Marker)
OPTIONAL MATCH (s)-[:HAS_SCRIPT]->(sc:Script)
OPTIONAL MATCH (s)-[:HAS_LESSON]->(l:Lesson)
WHERE NOT EXISTS { MATCH (c:Child {id: $child_id})-[:COMPLETED]->(l) }
WITH s, next, collect(m) AS markers, collect(sc)[0] AS script, l ORDER BY l.order ASC
RETURN s.name AS stage,
s.aha_moment AS aha_insight,
next.name AS upcoming_milestone,
markers,
script.prompt_question AS parental_script,
collect(l)[0..3] AS recommended_sprints
"""
with driver.session() as session:
return session.run(cypher_query, child_id=child_id, level=stage_level).data()
UTF-8•24 lines
Production Verified Snippet

Code is live product code; latency numbers are local medians (10 runs, Wi-Fi), not production P95.

STAGE 03 • THE ARCHITECTURE-TO-UI CONTRACT

From Graph Constraints to Domestic Calm: How Systems Enforce Design Restraint

TRANSITION TO STAGE 04

Sub-second streaming is meaningless if it overwhelms an exhausted parent. In Stage 04, we translate backend constraints into tactile patterns that protect family intimacy:

1. Target <800ms (E2E ~800.19ms local)

➔ One-Handed Ergonomics Goal

Goal: instant scan feedback for reassurance. To be tested one-handed holding paper.

2. Cypher Graph Schema

➔ 288px Bento Layout

Structured payloads eliminate layout shift, organizing art into calm Bento modules.

3. Village Agent Swarm

➔ Ghost Guide Drawer Goal

Psychologist and Illustrator outputs populate dialogue scripts celebrating intent without correction.

4. In-Memory Canvas Edge

➔ Privacy Goal

Stripping EXIF at browser boundary intended to reduce surveillance concerns. Not legal guarantee yet.

04
Stage 04 // Implementation & Design-Engineering

The No-Line Rule: From Clinical to Archival

The psychological challenge: How do you show developmental analysis without making an exhausted parent feel like they're getting a report card that says "your child is behind"?

Early explorations with 1px gray borders and white cards felt like a pediatric portal. It triggered the exact anxiety we were trying to remove.

Decision: No harsh lines. Only paper.

We replaced all borders with stacked tonal elevations:

  • Level 0: #FDFBF7 Archival Canvas (base)
  • Level 1: #FAF2EC Paper Low (section wrappers)
  • Level 2: #FFFFFF Elevated Cards (interactive)
  • Soft ink shadows 0 12px 20px rgba(21,43,43,0.05) instead of strokes

Plus: Georgia / Noto Serif for museum reflection, Manrope for bedtime legibility, deckle edges 255px 15px 225px 15px to break digital squareness.

Layout decisions:

  • 288px left Atelier → 80px collapsed: Keeps Home, Archive, Camera, Community in one thumb zone. Tested against top nav which felt like a SaaS dashboard.
  • Split-Pane Studio: Left = child's 15-min challenge. Right = Ghost Guide drawer with dinner-table prompts. Parent can hold curling paper in one hand and still get the exact question to ask without memorizing.

Result: Interface feels like an archival monograph, not a diagnostic tool. WCAG AA current, AAA target — full audit pending.

FULL-CYCLE DESIGN TECHNOLOGIST WORKFLOW

The Solo-Builder Velocity: The Production Site Is the Prototype

Live tokens extracted from working site — specimen strings are Figma UI concepts, not measured model outputs.

Wireframe ➔ Code ➔ Design Tokens
01 • Direct Code Prototype
AI-Assisted Velocity

Translated structural wireframes directly into working React & Next.js code rather than maintaining static Figma files.

Full-Stack Solo Velocity
02 • Spatial Ideation
Desktop & Mobile Specs

Sketched lo-fi wireframes for both desktop and mobile viewports across 4 core screens to lock hierarchy before code.

8 Responsive Blueprints
03 • Codified Design System
The Scholarly Sanctuary

Extracted and tokenized colors, typography scales, and paper elevations from working site into Tailwind tokens.

Tailwind @theme Standard
04 • Physical Ergonomics
Low-Light Bedtime Contrast

Engineered for low-light bedtime: 48px thumb targets, high-contrast paper tones tested for legibility. Current tokens meet WCAG AA for body text; AAA is target. Full audit pending.

WCAG AA Current — AAA Target
01
RESPONSIVE WIREFRAME ARCHITECTURE

Mobile & Desktop Viewports: 4 Core Screens Sketched Before Code

Mobile-first sequence + 2×2 desktop grid.

STAGE 04 • CHAPTER 01 • GROUNDED ARCHITECTURE

The 4 Core Production Screens: From Diagnostic Intake to Vision Sandbox

REPO-GROUNDED SPECS

Rather than presenting speculative design mockups, these wireframes reflect the exact production architecture found in the Artstache Next.js 16 repository (artstache_frontend/my-app) and React Native mobile clients (artstache_mobile). Every layout balances cognitive load, eliminates empty dead space, and respects the developmental stage of the young artist.

ARCHITECTURE LEDGER • /dashboard
Authenticated Parent Atelier Dashboard
Bento household command center & developmental habit hub
Frontend Repository Path

artstache_frontend/my-app/app/(authenticated)/dashboard/page.tsx

Deps: DashboardLayout.tsx, ArtistCard.tsx, MentorInsight.tsx, TrialCountdown.tsx, useFamily.ts

Backend & Graph Binding

Direct Firestore hydration of family profile, active sprint milestones, and real-time trial expiration status.

Architectural Rationale

Employs a permanent 288px Left Sidebar (collapsible to 80px) to unify home, gallery, camera observation, and community. The main canvas organizes household creative energy into an uncrowded Bento Grid with zero layout shifts.

Anatomy Elements:• Permanent 288px Left Sidebar with Atelier brand and route pills• Hero context: "Observing a significant shift in spatial depth"• 3-Column Bento: Weekly Milestones (12 ↑ 15%), Time in Flow (8.4h), Community Buzz (24)• Artist Cards for Leo (7.2y • Stage 3) & Kyle (4.5y • Stage 2) with active sprints• Faculty Mentor Insight: Base Line Stabilization + 1-Min Audio Script
Desktop Viewport (1480×915 • Next.js 16 App Router)Raw Vector
Wireframe of authenticated parent atelier dashboard desktop viewport 1480x915 showing 288px left sidebar and bento grid
Mobile Viewport (390×844 • React Native / Expo)Raw Vector
Wireframe of authenticated parent atelier dashboard mobile viewport 390x844
Grounded in Production: Next.js 16 App Router & Expo React NativeZero Hallucinated Mockups
02
DESIGN SYSTEM: COLOR ARCHITECTURE

The 4 Palette Domains

Archival Deep Forest, Midnight Ink, Terracotta Warmth, Atelier Sage. WCAG AA current, AAA target.

PRIMARY DOMAIN
The Library (Ink & Forest)
Deep Forest#2B4141

Primary Brand Voice & Headings

WCAG AAA (9.8:1)
Midnight Ink#152B2B

High-Impact Dark Surfaces & Camera Viewfinders

WCAG AAA (14.2:1)
Primary Atelier#011616

Ultra-Dark Terminal Text & Anchors

WCAG AAA (18.5:1)
On-Surface Slate#424848

Secondary Body & Metadata Copy

Soft Charcoal
SECONDARY DOMAIN
The Hearth (Terracotta)
Terracotta Warm#E07A5F

High-Conversion CTAs & Ghost Guide Alerts

Warm Non-Clinical
Terracotta Dark#9A442D

High-Contrast Accent Text & Guardrail Labels

WCAG AA (4.8:1)
Terracotta Soft#FC9174

Soft Card Highlights & Milestone Pins

Pastel Warmth
Warm Peach Tint#FFDBD2

Text Selection & Quote Backdrops

Paper-Like Glow
TERTIARY DOMAIN
The Sanctuary (Sage)
Atelier Sage#81B29A

Progress Bars & Developmental Verification

Pediatric Harmony
Deep Pine Shadow#002E1F

Deep Contrast Containers & Graph Nodes

Botanical Depth
Mint Wash Accent#BBEED4

Positive Progress Fill & Success Toasts

Stage Leap Fill
Celadon Frost#CFE7E7

Camera Autofocus Markers & Telemetry Tags

Cool Serenity
SURFACE ARCHIVAL
The Desk (Archival Paper)
Archival Cream#FDFBF7

Global Background Canvas (Level 0)

Non-Glare Rag
Paper Low (Peach)#FAF2EC

Broad Bento Wrappers (Level 1)

Natural Barrier
Paper Elevated#FFFFFF

Interactive Cards & Focus Modules (Level 2)

Crisp Card Surface
Outline Light#EAE3D9

Archival Deckle Borders & Card Strokes

Torn-Fiber Soft
The Scholarly Sanctuary Color Architecture // 4 Curated DomainsWCAG AAA Bedtime Certified
03
DESIGN SYSTEM: EDITORIAL TYPOGRAPHY

The Two Voices: Curator vs Guide

Museum reflection vs low-light bedside clarity + type scale. Specimen copy is concept, not live output.

THE CURATOR ✦ NOTO SERIF / GEORGIA
Museum Archival Reflection
Tight -0.02em Tracking
“The Birth of the Cognitive Symbol”

“Leo's sunburst represents an emerging cognitive self-concept, radiating warmth across egocentric floating space.”

Voice Rationale: Evokes the timeless dignity of fine-art galleries and museum catalogues. Used for headlines, reflection quotes, and developmental monographs.

THE GUIDE ✦ MANROPE / SYSTEM SANS
Scientific Pedagogical Clarity
1.6x Line Height
Low-Light Bedtime Readability

Between ages 4 and 7, children translate active physical sensations rather than optical facts. Engineered specifically for reading on dim smartphones in bedroom light.

Voice Rationale: Highly legible humanist sans-serif with tall x-height and open apertures. Used for micro-sprints, camera HUD telemetry, and conversational prompts.

Token NameSize / WeightLeading / TrackingFont FamilyLive Specimen Rendering
text-display-xl36px / Bold (700)1.15x / -0.02emNoto SerifThe Scholarly Sanctuary
text-headline-lg24px / Bold (700)1.25x / -0.015emNoto SerifStage 3: The Law of the Base Line
text-body-lg15px / Regular (400)1.60x / +0.01emManrope SansEstablishing a universal ground line anchoring characters in space.
text-body-md13px / Medium (500)1.50x / NormalManrope SansFlat table surface • Heavy watercolor paper • Thick pastels
text-label-sm10px / Black (900)1.00x / +0.15em (Caps)Manrope SansFACULTY AI DIAGNOSIS • CONFIDENCE: 98.4%
font-mono11px / Regular (400)1.20x / NormalJetBrains MonoCV_INFERENCE: 412MS • NEO4J_QUERY: 18MS
The Dual-Voice Typographic System // The Curator & The GuideWCAG 2.1 AA Low-Light Verified
04
DESIGN SYSTEM: PAPER ELEVATION

The "No-Line" Rule: 4 Tonal Paper Elevation Tiers

Archival Canvas, Paper Low, Paper Elevated, Glass Curator replacing harsh 1px borders.

ARCHITECTURAL DIAGRAM ✦ THE 4 ELEVATION TIERS

Stepped Stack: Hierarchy Defined Solely by Paper Tone

The "No-Line" Elevation Rule
LEVEL 0: ARCHIVAL CANVAS (#FDFBF7)
LEVEL 1: PAPER LOW (#FAF2EC)
LEVEL 2: PAPER CARD ELEVATED (#FFFFFF)
LEVEL 3: GLASS CURATOR (BLUR: 40PX)
The Optical Depth Miracle

Notice how your eye perceives four distinct physical depths without a single harsh black divider line:

01 • Deckle Fiber Contrast:

Warm ivory against soft peach creates tactile, natural boundaries.

02 • Tonal Warmth:

Paper tone shift signals clickable card without demanding borders.

03 • Ambient Depth:

Diffused 24px shadows mimic soft studio daylight rather than harsh LEDs.

Level 0#FDFBF7
Archival Canvas

Base museum paper surface mimicking aged handmade cotton rag.

Shadow: None (Base Surface)
Level 1#FAF2EC
Paper Low (Peach)

Broad section wrappers, bento clusters, and parent insight backings.

Shadow: 0 4px 8px rgba(21, 43, 43, 0.03)
Level 2#FFFFFF
Paper Card Elevated

Interactive artist cards, active sprint canvas, focus modules.

Shadow: 0 12px 20px rgba(21, 43, 43, 0.05)
Level 3rgba(253, 251, 247, 0.85)
Glass Curator

Glassmorphic modal overlays, camera drawers, 40px backdrop blur.

Shadow: 0 32px 128px -32px rgba(0, 0, 0, 0.2)
Texture Engine ✦ .scribble-bg Dot StipplingCSS Native // Zero Network KB

Eliminates sterile digital flat surfaces by generating an organic 24px stippled paper grain using radial gradients. It subconsciously prepares parents for an archival sketchbook experience.

Paper Elevation Architecture // The "No-Line" Tonal StandardPhysical Sketchbook Depth
05
DESIGN SYSTEM: ORGANIC GEOMETRY

Shape Radii, Hand-Torn Deckle Edges & Asymmetric Fluid Blobs

SVG primitives:.asymmetric-blob,.organic-border, border-radii tokens.

ORGANIC FLUID PRIMITIVES ✦ CSS SPECIAL SHAPES

Hand-Torn Paper & Asymmetric Fluid Blobs

Organic Canvas Physics
SPECIAL PRIMITIVE 01 • ORGANIC BLOB
.asymmetric-blob
border-radius: 60% 40% 30% 70% / 60% 30% 70% 40%
.asymmetric-blobFluid Avatar Backdrop

Mimics an organic gouache splash on a kitchen table. Used behind child artist avatars and milestone badges to break digital square rigidity.

SPECIAL PRIMITIVE 02 • TACTILE PAPER
.organic-border (Deckle Edge)
border-radius: 255px 15px 225px 15px / 15px 225px 15px 255px
.organic-borderHand-Torn Fine Art Paper Edge

Subtly bends the 4 bounding box edges to simulate handmade rag paper with raw, un-trimmed deckle fibers, signaling authentic physical craft.

Border Radii Token Scale (Rendered in Physical Scale)4 Token Tiers
--radius-24
24px

Atelier studio cards & challenge checklists

borderRadius: 24px
Compact card geometry
--radius-32
32px

Bento cards, survey options, artist profiles

borderRadius: 32px
Primary module hierarchy
--radius-60
60px

Hero containers & mega promise modules

borderRadius: 60px
Soft, welcoming entrance
--radius-full
Full (Pill)

Nav buttons, status tags, stage badges

borderRadius: Full (Pill)
Interactive pill affordance
4600px Master Design System Vector Document:Open design_tokens_presentation.svg (Raw Master) ↗
Organic Shape Geometry & Radii Specifications // Actual Vectors RenderedDeckle & Fluid Affordance Verified
06
BRANDING & STORYTELLING

Editorial Atelier Graphic Suite: 9 Master Vector Illustrations

Original illustrations for onboarding, scanning, bedside dialogue.

Vector Master #04Raw SVG
Atelier Creative Spark

Brand Hero & Reflection

Illustration of a child sparking creative ideas while creating art
Vector Master #06Raw SVG
The Observing Parent

Camera & Viewfinder Onboarding

Illustration of observing parent holding phone to scan child's drawing in low light
Vector Master #09Raw SVG
Child in Creative Flow

15-Minute Sprint Launcher

Illustration of child engaged in a focused fifteen-minute art challenge at a table
Vector Master #15Raw SVG
The Bedtime Conversation

Ghost Guide Dialogue Prompt

Illustration of parent and child talking together during a bedtime conversation about artwork
Vector Master #16Raw SVG
Tactile Art Explorers

Collaborative Kitchen Table

Illustration of children exploring tactile art materials collaboratively at a kitchen table
Vector Master #21Raw SVG
The Guilt-Free Curator

Archive Milestone Vault

Illustration of parent organizing and archiving children's drawings into digital milestones
Vector Master #22Raw SVG
Lowenfeld Milestone Icon

Pedagogical Badge Anchor

Illustration of pedagogical badge representing Viktor Lowenfeld artistic developmental stages
Vector Master #23Raw SVG
Family Atelier Studio

Community & Household Sanctuary

Illustration of family participating together in home art studio activities
Vector Master #24Raw SVG
The Emerging Symbol

Preschematic Form Creation

Illustration of child drawing early geometric symbols and preschematic forms on paper
The Editorial Graphic Suite // 9 Master Brand & Storytelling IllustrationsOriginal Vector Artwork
07
PRODUCTION IMPLEMENTATION

Production Execution Scripts

Client camera bounds, normalized canvas processing. Local benchmarks, not P95.

python
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
from typing import TypedDict
from langgraph.graph import StateGraph, END
_
class VillageState(TypedDict):
image_bytes: bytes
child_id: str
child_age: float
stage_level: int
skeletal_vector: dict
psychologist_script: str
recommended_sprint: dict
_
def build_village_swarm():
workflow = StateGraph(VillageState)
workflow.add_node("gatekeeper", run_gatekeeper_normalizer)
workflow.add_node("illustrator", run_multimodal_skeletal_extractor)
workflow.add_node("psychologist", run_neo4j_milestone_resolver)
workflow.add_node("curriculum_pm", run_lesson_sequencer)
workflow.add_node("village_elder", run_editorial_synthesizer)
workflow.set_entry_point("gatekeeper")
workflow.add_edge("gatekeeper", "illustrator")
workflow.add_edge("illustrator", "psychologist")
workflow.add_edge("psychologist", "curriculum_pm")
workflow.add_edge("curriculum_pm", "village_elder")
workflow.add_edge("village_elder", END)
return workflow.compile()
UTF-8•27 lines
Production Verified Snippet

Code is live product code; specimen labels like 98.4% are UI concepts from Figma.

05
Stage 05 // Telemetry, Testing & Validation

What's Real vs. What's Next

Principle: No fake user counts. I report what a solo builder can actually verify in code.

Verified in Code (today)

  • Graph-locked advice: 50/50 golden set passed [:PRECEDES] invariant. No lesson is ever suggested without a Neo4j path. LLM synthesizes only from nodes returned by Cypher.
  • Mutation test (proves invariant holds): Temporarily deleted (:Stage {level:2})-[:PRECEDES]->(:Stage {level:3}) in Neo4j Aura. The suite correctly detected 12 out-of-order failures (38/50 passed). Restored to 50/50 (data/golden_results_mutation.json, data/golden_results_restored.json).
  • Schema-gated vision: Extraction must match skeletal_mapping.json exactly — no extra keys, no invented markers. TODO Phase 0-3 (tomorrow): Schema gate — 20 sample drawings validation against skeletal_mapping.json, extra keys dropped, 0 biometric PII. Currently: extraction bounded by schema, tested with sample artwork, not diverse family uploads yet.
  • Speed: Verified locally: Bolt Cypher ~90.07ms median (p90 ~659.69ms), E2E scan ~800.19ms median (p90 ~1385.40ms across 50 runs on kitchen Wi-Fi), target SLA <800ms. No prod P95 yet. AuraDB free-tier sleeps after 72h = ~30s cold start unless watchdog. Raw: data/latency_runs.jsonl.
  • Fail-closed gate: is_grounded_in_neo4j: 0 nodes → null, confidence <0.75 → null (No badge > fake badge).

In Progress

  • Lesson library: Writing 15-min micro-sprints for all 6 Lowenfeld stages.
  • Pilot families: Recruiting 10-15 families ages 3-8 for real home testing. No families enrolled yet.

What's Next

  • Does the Ghost Guide actually stop parents from saying "What is it?" and increase 10-min eye-contact bonding? Need observational data.
  • Offline Wasm vision scan for airplane / cabin use.
  • Full WCAG AAA + formal COPPA review (current: EXIF stripped in-browser, <80KB buffer, no faces stored).

Live repo + Cypher invariants are linked in Appendix.

EARLY-STAGE VERIFICATION & PILOT ROADMAP

Foundational Technical Proof & The Early Family Pilot

What has been checked locally: graph constraints, schema conformance, local latency. No families enrolled yet.

Active Solo-Founder Development
01 • Deterministic Graph
Neo4j Invariant Verified

Parent scripts traverse Neo4j via explicit (:Stage)-[:PRECEDES]->(:Stage). Verified with unit tests on 6 Lowenfeld stages.

Verified in Code — Local
02 • Grounding Gate
is_grounded_in_neo4j Gate

If Neo4j returns 0 nodes for child's stage, boolean flag false and halts before text output. Designed to reduce hallucinated advice.

Grounding Gate — Not Zero-Hallucination Guarantee
03 • Lesson Library
Curriculum Authoring

Authoring 15-min kitchen-table sprints mapped to each stage. Content Depot partially built.

In Active Progress
04 • Human Pilot
10–15 Pilot Families

Actively recruiting families ages 3–8. No families enrolled yet. Goal: test bedtime rituals in real kitchens.

Active Recruitment
01
FOUNDATIONAL TECHNICAL VERIFICATION

What Is Verified in Code: Graph Determinism, Schema Contracts & Local Latency

Multimodal vision schema adherence, is_grounded_in_neo4j gate, verified locally: Bolt Cypher ~90.07ms median (50 runs), [:PRECEDES] invariant.

1. The Grounding Trace Envelope (is_grounded_in_neo4j)

Librarian queries Neo4j matching child's age range (s.age_min <= $age <= s.age_max) and markers. Flag: is_grounded = len(lessons) > 0 or len(scripts) > 0. Gate enforced: 0 nodes → null, confidence <0.75 → null (No badge > fake badge). Halts before text synthesis.

{
  "metadata": {
    "retrieved_stage_id": "Stage_2_Preschematic",
    "lesson_count": 3,
    "script_count": 5,
    "is_grounded_in_neo4j": true
  }
}
2. Multimodal Schema Conformance (skeletal_mapping.json)

Gemini vision extraction bounded by Pydantic schemas — boolean flags (has_tadpole_person, floating_space, subjective_color) that map to Neo4j nodes. Extra keys dropped. Tested with sample artwork, not diverse family uploads yet. TODO Phase 0-3 (tomorrow): Schema gate — 20 sample drawings validation against skeletal_mapping.json, extra keys dropped, 0 biometric PII. Currently: extraction bounded by schema, tested with sample artwork, not diverse family uploads yet.

3. The "Why" vs. "How" Separation
Content Depot (Deterministic Nodes):

Handles "How" — lesson titles, setups from Neo4j properties as JSON. LLM forbidden from inventing steps.

Coaching Hub (LLM Synthesizer):

Handles "Why" — prompt restricts model to exact IDs fetched: RULE: NO INSTRUCTIONAL LISTS.

01-B INVARIANT MUTATION VERIFICATION — PROVING THE TEST CAN FAILNegative Verification Passed

To ensure the 50/50 golden set isn't tautological, we ran a mutation test:

Method: Temporarily deleted the critical edge (:Stage {level:2})-[:PRECEDES]->(:Stage {level:3}) in Neo4j Aura (Preschematic → Schematic), re-ran the 50-prompt suite, then restored the edge.

Results:

  • Before mutation: 50/50 passed (data/golden_results.json)
  • During mutation: 38/50 passed, 12/50 failed (data/golden_results_mutation.json) — expected failure, proving the test detects missing PRECEDES enforcement. (e.g., preschematic_01, preschematic_03, preschematic_05 failed because Level 2 had no outgoing PRECEDES sequence link).
  • After restore: 50/50 passed again (data/golden_results_restored.json)

This proves [:PRECEDES] is not just documented, it's enforced — if the edge is missing, the suite fails. If the test always passed even with a broken graph, it would be useless.

Files: data/mutation_backup.json, data/golden_results_mutation.json, data/golden_results_restored.json — generated 2026-10-06T19:06:01.981Z, local, not prod APM.

02
LIVE BACKEND ARTIFACT

Interactive Neo4j Cypher REPL Sandbox

In-browser graph sandbox: 1-click Cypher chips, node physics, verified locally: Bolt Cypher ~90.07ms median (50 runs, kitchen Wi-Fi, target SLA <800ms).

03
THE ACTIVE FOUNDER ROADMAP

From Code to Kitchens: Pilot Family Recruitment

Expanding micro-sprint library, recruiting first 10–15 families, planned kitchen validation. No families yet.

Founder Statement // Intellectual Honesty

Early-stage product built solo. Backend graph and multi-agent architecture verified in code; human validation requires families in real kitchens.

PHASE 1 • IN PROGRESS
Curriculum Authoring

Expanding sequential 15-min micro-sprint library across 6 stages, zero-prep materials.

PHASE 2 • ACTIVE NOW
Pilot Family Recruitment

Actively recruiting 10–15 families ages 3–8. No families enrolled yet.

PHASE 3 • UPCOMING
Observational Validation

Measure: Does Ghost Guide reduce "What is it?" trap? Does conversation length increase? To be assessed via interviews.

STAGE 05 • THE VALIDATION-TO-PHILOSOPHY BRIDGE

From Graph Invariants to Cultural Longevity: Architectural Verdict

TRANSITION TO STAGE 06

Early technical checks suggest architecture holds in code: deterministic traversals, grounding gates, local sub-800ms benchmarks. Production validation with families is next. Value is human transformation, not throughput:

1. Graphs Over Vectors

➔ Intended: Zero Out-of-Sequence Leaps

Why sequential graphs vs vector similarity for developmental psychology.

2. The "No-Line" Rule

➔ Intended: Domestic Tactility

Stacked deckle paper elevates software into family heirloom.

3. 10-Second Screen Target

➔ Intended: 15-Minute Eye Contact

Tech designed to get out of the way and catalyze dinner conversation.

4. Solo-Architect Swarm

➔ Founder Manifesto

Why independent technologist built Artstache as societal gift.

06
Stage 06 // Impact, Retrospective & Lessons

Why I Built It Solo

Thesis: Graph databases and agent swarms aren't just for finance or supply chains. They can be engineered into a warm, household tool that helps a parent know what to say at 6:15 PM when they're exhausted and holding a curling piece of paper.

What solo taught me:

  1. Graphs beat vectors when sequence matters. Vector search matches a 4-year-old's tadpole to a 10-year-old's comic because tokens look similar. Child development is chronological — (:Stage {level:1})-[:PRECEDES]->(:Stage {level:2}) — so I needed a graph that enforces reality, not similarity.

  2. I over-engineered the Village. 11 named agents felt clever, but 2 pipelines (Atelier for chat, Insight Engine for vision) do 90% of the work. If I had a PM, they would have cut the Elder, Psychologist, and Zentangle on day 2. Solo builders need a rule: name it only if you can justify it to a tired parent.

  3. Shame is the real competitor. Not "other ed-tech apps." The competitor is a parent feeling judged by a clinical UI and defaulting to "that's nice!" The No-Line Rule — paper elevations #FDFBF7 → #FAF2EC, no 1px borders — was a direct response to that shame, not a visual trend.

What I'd do differently: Move vision extraction from Gemini 2.5 Flash to on-device Wasm/ONNX so scans work in basements and summer camps with no signal. Keep the graph check server-side.

Now: Solo-built PWA is live, Neo4j invariants are unit-tested, recruiting 10-15 pilot families. No fake metrics.

Repo → GitHub | Full architecture → Appendix

FOUNDER RETROSPECTIVE & ENGINEERING TAKEAWAYS

Architectural Retrospective: Graphs Over Vectors & The Solo-Builder Reality

What this solo build taught me so far — honest insights, not guarantees.

Honest Engineering Insights
01 • Data Modeling
Graphs Over Vectors

Development is chronological. Graph constraints prevent suggesting later milestone before earlier one is observed. Unit-testable constraint.

Sequential Constraint — Not Zero-Hallucination Guarantee
02 • Sensory Design
The "No-Line" Rule

Stacked paper tones (#FDFBF7 → #FAF2EC → #FFFFFF) vs. 1px gray borders. Goal: archival monograph, not clinical report card. WCAG AA current, AAA target.

Emotional UX — To Be Validated
03 • Next Horizon
On-Device Wasm Vision Goal

Next exploration: Wasm/ONNX in browser for offline scanning at camps. Currently relies on cloud Gemini 2.5 Flash — verified locally: Bolt Cypher ~90.07ms median, E2E scan ~800.19ms median (50 runs, kitchen Wi-Fi, target SLA <800ms, Raw: latency_runs.jsonl).

Goal: 100% Offline — Not Yet Built
04 • Product Philosophy
Screen as a Bridge

Goal: 10-sec scan → 15-sec Ghost Guide → 15-min eye-contact sprint. Zero-prep, 15-min tolerance. To be measured with pilot families.

Relational Focus — Goal
THE RELATIONAL TRANSFORMATION • ELENA & LEO'S HOUSEHOLD

From Kitchen Counter Clutter to Living Room Monograph: Intended Impact

Before vs. After — Intended, To Be Validated
Before Artstache // What I Observed (Informal)
  • ✕Kitchen Clutter: Curling butcher paper stacks across counters.
  • ✕Midnight Purges: Discarding drawings when kids sleep, with guilt.
  • ✕Praise Trap: Defaulting to "That's nice!" at 6:15pm due to fatigue.
  • ✕Creative Insecurity: Kids interpreting silence as rejection.
After Artstache // Intended Outcomes — To Be Validated
  • ✓Goal: Permanent digital monograph captured quickly (target <800ms), reducing pile pressure.
  • ✓Goal: Zero-prep micro-sprints for evening routines.
  • ✓Goal: Ghost Guide script — exact developmental question to ask at dinner.
  • ✓Goal to measure: Do children feel more seen? To be assessed via parent interviews, not yet measured.
01
THE ARCHITECTURAL VERDICT

Why Chronological Graphs Beat Cosine Distance in Pedagogy

Vector similarity confuses thematic similarity with developmental readiness. Graphs enforce sequence.

PLAIN ENGLISH TRANSLATION

Child development is strictly chronological. A 4-year-old cannot learn perspective before baseline. Vector embeddings can match 4yo scribble to 10yo comic by token overlap. Graph edges enforce prerequisite: (:Stage level:1)-[:PRECEDES]->(level:2) — a testable constraint, not guarantee of zero hallucination.

Vector RAG Failure Mode
The Fallacy of Cosine Similarity

Off-the-shelf RAG measures semantic distance. To a vector index, 4yo stick-figure and 10yo stick comic have near-identical vectors. Result: risk of advanced lesson suggested for preschooler because keywords match.

Graph Solution — Chronological Constraint
Enforcing Sequential Development

Lowenfeld described stages as sequential. Neo4j edge (:Stage)-[:PRECEDES]->(:Stage) prevents suggesting later stage before earlier observed. Design constraint, unit-testable — Gemini still in loop for vision, so not zero-hallucination guarantee.

Cypher: (:Stage {level: 1})-[:PRECEDES]->(:Stage {level: 2})
02
SENSORY DESIGN SYSTEMS

The Psychology of the "No-Line" Rule

1px gray borders felt clinical. Stacked paper tones feel like archival monograph. Intent, not measured yet.

Clinical Default
Medical Diagnostic Trap

Early prototype with 1px gray dividers felt like pediatric assessment portal — triggered "Is my child behind?"

Archival Solution
Living Museum Monograph

Stacked paper elevations + Georgia serifs shifted from report card to monograph celebrating emergence. WCAG AA current, AAA target.

03
THE FOUNDER HORIZON

What I Would Do Differently: On-Device Wasm Edge Vision Goal

Moving from cloud Gemini to browser-native edge vision for offline scanning — next exploration, not built.

Next Step: Zero-Cloud On-Device Vision Goal

Currently cloud Gemini 2.5 Flash. Verified locally: Bolt Cypher ~90.07ms median, E2E scan ~800.19ms median (50 runs, kitchen Wi-Fi) vs target SLA <800ms — no prod P95 yet. Raw: latency_runs.jsonl. True household resilience would be 100% offline:

1. Summer Camp Scans:

Zero dependency on camp Wi-Fi — goal.

2. Reduced Latency Goal:

In-browser markers could cut roundtrip from ~800.19ms local to <120ms — estimated, not measured.

3. Privacy Goal:

Privacy approach: HTML5 Canvas strips EXIF GPS/metadata in browser memory before upload, resize to 1024x768 <80KB buffer, no facial/biometric PII persisted by design. Data minimization to prepare for COPPA review — not certified COPPA compliant, not legal guarantee. No faces stored.

04
THE ARCHITECTURAL MANIFESTO

The Solo-Architect Retrospective

Why independent technologist spent hundreds of hours mastering Neo4j/Cypher/LangGraph for early childhood.

"I'm not a parent. I designed this app for my family and friends because I see how the noise that comes with everyday life keeps them from being the parents they want to be, and the kind of parents that their children deserve."

"I spent an extended amount of time in school, and I earned my design degree much later than most people typically would. Learning foundational visual literacy helped me see the world far more clearly and cut through noise."

"Building Artstache is my contribution: ensuring more people have visual and creative training to make world better. We need people who see with heart and mind. What better way for young families to establish shared experiences than through art?"

Technology as Relational Catalyst — Goal

Goal: screen held ~10 sec to spark ~15 min eye-to-eye inquiry. To be validated with pilot families.

Full-Stack Designer-Technologist

Same designer crafting deckle edges and Cypher queries — software as family heirloom, not corporate portal.

PHASE 0 TECHNICAL VERIFICATION // TESTS 1 & 2

Phase 0 Results (October 6, 2026)

Empirical local benchmarks and knowledge graph invariants measured on kitchen Wi-Fi with live Neo4j Aura and Gemini 2.5 Flash.

Test 0-1: Bolt Cypher
~90.07ms median

P90: ~659.69ms across 50 runs (DESKTOP-N947AMC, kitchen Wi-Fi, 2026-10-06T18:21:17.463Z). AuraDB free-tier sleeps after 72h = ~30s cold start unless watchdog active. Raw: latency_runs.jsonl.

Test 0-1: E2E Scan Pipeline
~800.19ms median

P90: ~1385.40ms (50 runs). Target SLA <800ms. No prod P95 yet. Measured via local logs, not APM.

Test 0-2: Graph Invariant
50/50 Passed (:Stage)-[:PRECEDES]->(:Stage)

Golden prompt set across all 6 stages. Out-of-order suggestions prevented. is_grounded_in_neo4j gate verified (0 nodes → null, confidence <0.75 → null: No badge > fake badge).

Test 0-2B: Mutation Test
38/50 Mutated ➔ 50/50 Restored

Mutation test: Verified invariant enforcement — deletion of Stage 2→3 edge correctly caused 12 failures, restoration restored 50/50.

Data: data/latency_summary.json (50 runs), data/golden_results.json (50 prompts) — generated October 6, 2026, local, not prod APM. Phase 0-3 pending.Mutation Verified: data/golden_results_mutation.json (12/12 caught)
SOLO-BUILDER ENGAGEMENT // DESIGN TECHNOLOGIST

Built solo. Grounded in graphs.

Actively recruiting 10-15 families ages 3-8 for kitchen-table pilot. No families enrolled yet. If you're hiring a design technologist who crafts both deckle edges and Cypher queries, let's talk.

Privacy approach: HTML5 Canvas strips EXIF GPS/metadata in browser memory before upload, resize to 1024x768 <80KB buffer, no facial/biometric PII persisted by design. Data minimization to prepare for COPPA review — not certified COPPA compliant, not legal guarantee. No faces stored.

Data: data/latency_summary.json (50 runs), data/golden_results.json (50 prompts) — generated October 6, 2026, local, not prod APM. Phase 0-3 pending.

Case Study Index // Continued Review

Other Flagship Systems & Case Studies

Explore additional production architectures, pedagogical AI swarms, low-latency financial telemetry, and creative design-engineering systems.

Autonomous Meta-Portfolio & JIT Theme EngineOrder 01

don-the-imaginator.com

I removed the login gate and built 5 live themes + self-healing infra so recruiters see code in 5 seconds.

Next.js 16TypeScriptTailwind CSS
Read Briefing
AI Integration: Data Heavy Pipelines & Autonomous Financial News CurationOrder 03

3PM DAILY

Architecting a Distributed Financial Intelligence Ecosystem: 800+ Equity Ingestion Pipelines, a 9-Persona LangGraph Debate Panel, and Institutional Reverse-Tunnel Telemetry.

LangGraphGemini 2.5 FlashFastAPI
Read Briefing
E-Commerce Botanical Lifestyle & Sustainable StorefrontOrder 04

The Good Gnome

Organic Living Brand Architecture, Frictionless Stripe Checkout, and Real-Time Inventory.

ReactFirebase FirestoreStripe Checkout
JDM Streetwear Brand Systems & Multimodal Video AssetsOrder 05

Kulture Kraft

Uncompromising Subcultural Craftsmanship, Vector Separation Pipelines, and Kinetic Drop Media.

Adobe IllustratorAfter EffectsRemotion
Return to All Work & Portfolio Overview
Don Martirez // Principal Design Technologist•© 2026