Echology Machine Map

Every function, connection, strategy, and task mapped as a single system. The factory floor of echology.io. Generated 2026-03-15.

11Projects
~40KLines of Code
764Tests
84API Routes
30+Agent Skills
14DB Tables
8Docker Services
2Machines

Architecture

Layer Stack — Dependency Chain

Everything flows one direction. No layer depends on anything above it. decompose depends on nothing.

Foundation (L0)
Platform (L1)
Operations (L2)
Verticals (L3)
External / Personal
DECOMPOSE L0 — Deterministic Foundation — PyPI — 4,837 LOC — 73 tests — Zero deps classify() · decompose_text() · filter_for_llm() · 6 authority · 7 risk · MCP server ENGINE L1 — Shared Intelligence — ~15,000 LOC — 19 domain seams — 70% universal VANTA (Perceive) ALETHEIA (Verify) DAEDALUS (Act) OPS L2 — Command Center — Dashboard · Agent · 14 tables · 30+ skills · Training · CLI CRM · Financial · Content · Corpus · Scoring · Email · Invoices · Citations · Calendar AECai AEC Vertical — v3.1.0 84 routes · 6 workflows 569 tests · Port 8443 FastAPI + Temporal + Qdrant GEODE GEO Vertical — Phase 1 14 pages (EN+PT) 16 queries · 5 platforms Citation CLI · geode.digital RBS-DEMO Insurance Vertical 2,150 LOC · 40+ fields 3-tier extraction Regex → Decompose → LLaVA OPEN-SCRIPTURE Scripture Dataset 31,102 verses · 344K edges 9,971 chunks · KG 101MB WEB + KJV translations PRISM KYLEVINES LAB INFRASTRUCTURE — Docker (Qdrant · Temporal · Caddy) — Ollama (6 models) — Surface Laptop (dev) — Mac Mini (ML/prod) — GitHub Pages (sites) — GHCR (images)

Pipeline

INGEST → CLASSIFY → EMBED → RETRIEVE → ACT

Every document passes through the same pipeline. Deterministic classification happens before any LLM touches the data.

INGEST Vanta Stage 1 PDF extraction (PyMuPDF) DOCX parsing DXF/CAD geometry OCR (Tesseract) Text chunking (2000/200) CLASSIFY decompose + Vanta Stage 2 DETERMINISTIC FIRST Authority (6 categories) Risk (7 categories) Entities + Irreducibility Attention scoring (0-10) EMBED Vanta Stage 3-4 nomic-embed-text (Ollama) sentence-transformers Qdrant vector upsert AI enrich (confidence-gated) Only if confidence < 0.80 RETRIEVE Daedalus + Aletheia Qdrant vector search Schema validation Jurisdiction cross-ref Audit ledger (SHA-256) Certification (gold/silver/bronze) ACT Daedalus + Human Gate Intelligence reports Civil3D / Revit scripts RAG chat responses Outreach drafts Training data export DETERMINISTIC — No LLM, no GPU, no hallucination GATED — LLM if needed VERIFIED — Provenance tracked, human-approved PDF · DOCX · DXF · TXT · MD CSV · HTML · JSON · Images

Operations

Ops Command Center

The nervous system. 14 database tables, 30+ agent skills, human-in-the-loop gates on every external action.

OPS DASHBOARD Port 8444 — FastAPI + HTMX 21 routes · 21 templates · Zero JS frameworks Every mutation is a proposal until approved HUMAN GATE ON ALL EXTERNAL ACTIONS SQLite — ops.db — 14 Tables merit_logDaily M.E.R.I.T. tracking leadsCRM pipeline + scoring interactionsEvery touchpoint logged financial_trackerRevenue + expenses tasksNotion sync + new tasks corpus_indexIngested documents corpus_chunksText chunks + embeddings decision_logAudit trail (reversible?) weekly_reviewSaturday review snapshots objectives"I Will Obtain" targets content_queueDraft → review → publish citation_queriesGEO query tracking citation_checksPlatform citation results engagementsClient engagement tracking + idea_archive, followup_log, onboarding, case_study_log, citation_surfaces, lead_sequences Agent Orchestrator — 30+ Skills EXECUTION LOOP 1. Read situation (structured state) 2. Reason (LLM analyzes priorities) 3. Execute through skills 4. Gate external actions 5. Log everything SKILL CATEGORIES Revenueleads_list, score, stage, stale Financialsummary, by_category, update Contentdraft, queue, publish Outreachdraft_email, approve, send Taskslist, create, update IntelSOP violations, forecast, reconcile MODULES context · executor · gateway · memory · prompts

Infrastructure

Docker Services + Hardware

Service

Qdrant (v1.13.2)

Vector storage for embeddings. HTTP + gRPC.

  • Port 6333 (HTTP), 6334 (gRPC)
  • 2 CPU, 2 GB RAM
  • Persistent volume: qdrant_data
Service

Temporal (v1.26.2)

Workflow orchestration for long-running tasks.

  • Port 7233 (gRPC)
  • UI at port 8233
  • SQLite backend
Service

Ollama

Local LLM inference. Runs on host GPU, not in container.

  • Port 11434
  • Models: llama3, llava, nomic-embed-text, aecai, phi3, gemma2
  • Primary: Mac Mini (Apple Silicon)
Service

Caddy (v2-alpine)

Reverse proxy with automatic TLS.

  • Ports 80, 443
  • echology.io → server:8443
  • Rate limiting: 60 req/min/host
Hardware

Surface Laptop 4

Primary development machine.

  • Pop!_OS 24.04, i7-1185G7, 16GB
  • LAN: 10.0.0.64
  • Docker, Claude Code, VS Code
Hardware

Mac Mini

ML inference + production services.

  • macOS 26.3 (Tahoe), Apple Silicon
  • LAN: 10.0.0.118
  • launchd: io.echology.aecai, io.echology.ops
  • cloudflared: aecai.io + ops.aecai.io

External

Public Surfaces + Deployments

Website

echology.io

Company marketing site. GitHub Pages.

  • Product pages, blog, decompose docs
  • Portuguese translations
  • GEO citation surfaces
Platform

aecai.io

AEC document intelligence. Cloudflare tunnel to Mac Mini.

  • Port 8443 via cloudflared
  • Public: demo, engage, GEO pages
  • Localhost: process, app, chat, outreach
Website

geode.digital

GEO services brand. GitHub Pages.

  • 14 pages (EN + PT)
  • Build log, methodology, contact form
  • Form → ops.aecai.io/api/leads/inbound
Website

kylevines.com

Personal site. GitHub Pages.

  • Blog, about, contact, projects
  • field.js wave canvas
  • CRT terminal aesthetic
Package

PyPI: decompose-mcp

Published Python package. v0.2.0.

  • pip install decompose-mcp
  • MCP server + CLI
  • MIT License
Registry

GHCR Images

Docker images built on push to main.

  • ghcr.io/echology-io/echology-server
  • ghcr.io/echology-io/ops-dashboard
  • GitHub Actions CI/CD

Strategy

Foundation Values + Structural Principles

FIVE CORE VALUES I. Truth over plausibility Deterministic classification before probabilistic generation II. Structure over volume Reduce noise. Increase density. Less that is true > more that is approximate III. Provenance over assertion Immutable audit trails. Origin determines nature. Chain never broken IV. Perception over generation See before speaking. Classify before inferring. Understand before responding V. Integrity under pressure Accuracy wins over speed, cost, market pressure, or impressiveness STRUCTURAL PRINCIPLES Unified System Not a feature collection. One coherent architecture. Semantic Proximity Information that shares meaning shares structure. Find by meaning. Workflow as Architecture Carry intelligence from source to decision-maker without distortion. AI as Lens Not answer machines. Lenses that sharpen perception of existing structure. Deterministic First Foundation must be reliable. Inference on inference is noise.

Training

Three-Layer Model Architecture

LAYER 1: echology:latest Platform Intelligence 3,319 training pairs (1,319 foundation + 2,000 supplemental) Doc target: 1,500-2,300 (exceeded) Sources: FOUNDATION, CLAUDE, architecture, ops, prompts Perplexity: 6.427 LAYER 2: aecai:latest Industry Knowledge 1,173 pairs (1,100 platform + 73 AEC) Target: 5,000+ pairs Sources: base + Vanta pipeline, domain docs, SOPs, marketing Perplexity: 5.415 LAYER 3: {customer}:latest Customer-Specific Their specs, contracts, RFIs Their meeting minutes, lessons Runs on their hardware NEVER LEAVES THEIR NETWORK MLX LoRA on Apple Silicon

Growth Engine

Geode Self-Validating Loop

Geode executes GEO methodology Portfolio becomes AI-cited echology + AECai + Geode Buyers find via AI search Methodology proven on portfolio Sell to external clients Case studies validate echology technology LOOP COMPOUNDS Baseline (2026-03-14): Geode 0/6 · echology 1/5 · AECai 2/5 — 90-day target: 5/6 · 4/5 · 5/5

Inventory

All Projects

L0 Foundation

decompose

Deterministic text classification. Zero deps. Published on PyPI. The thing everything depends on.

  • 4,837 LOC · 73 tests
  • 3 public functions: decompose_text, decompose, filter_for_llm
  • 6 authority categories, 7 risk categories
  • 10 irreducibility patterns
  • MCP server (2 tools) + CLI
  • Entity extraction: standards, dates, financial, legal refs
  • Semantic chunking (Markdown-aware, 2000/200)
  • Marketing automation pipeline
  • Performance: ~14ms per document, 1000+ chars/ms
L1 Platform

engine

Three-engine document intelligence. Domain-agnostic core with 19 seams for vertical configuration.

  • ~15,000 LOC · 70% universal
  • Vanta: 13 modules (core, classify, pipeline, ai, plugins, embed, index, batch, geometry, security, torsion)
  • Aletheia: 6 modules (schema, jurisdiction, ledger, cli)
  • Daedalus: 6 modules (retrieve, report, scripts)
  • Config: dual env var support (ECHOLOGY_* / AECAI_*)
  • Domain loader: ECHOLOGY_DOMAIN env var
  • Torsion: adaptive computation (cognitive systems 1-4)
  • 13 simulation systems consolidated
L2 Operations

ops

Command center. CRM, financial, content, corpus, agent orchestrator, training pipeline, dashboard.

  • ~4,200 LOC (core + dashboard)
  • 14 database tables, 40+ query functions
  • 30+ agent skills across 6 categories
  • Dashboard: 21 routes, 21 templates, HTMX
  • CLI: generate, ingest, sync, week, check, train, dashboard
  • Agent: context, executor, gateway, memory, skills, prompts
  • Lead scoring: ICP, persona, intent, engagement (deterministic)
  • Charts: radar, bar, matrix, waterfall, timeline (SVG)
  • Invoice generation, email (SMTP), sequences
  • Calendar: custody schedule, planetary days
  • Training pipeline: 3-layer model architecture
L3 Vertical

aecai

AEC document intelligence. FastAPI + Temporal + Qdrant. Full pipeline with verification and retrieval.

  • v3.1.0 · 569 tests · 84 routes
  • 6 Temporal workflows, 20+ activities
  • 4 domain modules: aec.py, civil3d.py, revit.py, geometry.py
  • TEA breach protocol (file security)
  • OCR: PyMuPDF + Tesseract
  • RAG chat (Qdrant + Ollama streaming)
  • Deliverables: 6 report templates
  • GEO surfaces (EN + PT)
  • Docker: decompose → engine → ops → aecai layer
L3 Vertical

geode

AI citation optimization (GEO). Dual function: portfolio marketing engine + external service.

  • Phase 1 active (Foundation)
  • 14 site pages (7 EN + 7 PT) on geode.digital
  • 16 baseline queries across 5 platforms
  • Citation CLI: query, check, report, surfaces, surface
  • Strategy: 784-line vertical strategy doc
  • Revenue target: $5-10K/month at 90 days
  • Self-validating loop: prove on portfolio, sell externally
  • 3 portfolio clients: echology, AECai, Geode
L3 Vertical

rbs-demo

Insurance policy QC. 3-tier extraction: regex, decompose+Ollama, LLaVA vision.

  • 2,150 LOC single-file server
  • 40+ extracted fields across property, GL, auto
  • 3-tier: native PDF regex → AI text → vision model
  • Compare quotes vs policies vs TAM exports
  • XLSX export for discrepancy reports
  • SOP + HITL touchpoint documentation
L3 Data

open-scripture

AI-ready Bible dataset + knowledge graph. Proves cross-domain thesis.

  • 31,102 verses · 344,799 edges
  • 9,971 passage chunks (retrieval-optimized)
  • Knowledge graph: 101MB (nodes + edges)
  • 3 data layers: canonical, chunks, graph
  • WEB + KJV translations (public domain)
  • Domain module for engine integration
Personal

prism

Desktop email/calendar concept. Tauri + React + Rust. Currently a stub.

  • Template boilerplate only (src/ empty)
  • ARCHITECTURE.md planned but unbuilt
  • No active development (>2 months idle)
  • Not core echology
R&D

lab

Test harness, simulations, research. Anthropic prompt corpus.

  • MCP test client (test_mcp.py)
  • 4 simulation systems: Vanta, Vanta torsion, Aletheia, Daedalus
  • Anthropic corpus: 10 prompting guides
  • Experimental ground before production
Marketing

echology-site

echology.io marketing. Product pages, blog, decompose docs, GEO surfaces.

  • GitHub Pages deployment
  • Product pages: aecai, decompose, geode
  • Blog with posts
  • Portuguese translations
  • GEO citation surfaces
  • Logo drafts (6 iterations)
Personal

kylevines

kylevines.com. Personal site, blog, field.js wave canvas.

  • GitHub Pages deployment
  • CRT terminal aesthetic
  • Blog: "Why I Build"
  • Projects: echology, AECai, decompose, Geode
  • Dark/light mode toggle
  • Logo drafts (v1, v2)

The Test

Does this reveal structure or introduce noise?

If it reveals structure, build it. If it introduces noise, don't.