Platform State: March 2026
What is built, what is running, and what is next. Three engines, three trained models, five applications, zero cloud dependencies.
Note: This build log was written when Surface was named Geode and Signal for AEC was named AECai. Model identifiers and application names reflect conditions at publication time.
Echology is a platform company. Three engines, trained models, and multiple applications all running on the same architecture. This post is a transparent accounting of what exists as of today.
The engines
The core of the platform is three engines that transform raw documents into structured, verified, actionable intelligence. They are domain-agnostic. Applications configure them for specific industries.
Vanta
Perceive. 5-stage pipeline: parse, classify, enrich, decompose, index. 13 modules. Handles 16+ file formats. Deterministic classification before any LLM is invoked.
Aletheia
Verify. Schema validation, jurisdiction cross-reference, SHA-256 hash-chained audit trail. Every classification is traceable and tamper-evident.
Daedalus
Act. Vector retrieval via Qdrant, intelligence report generation, CAD/BIM script generation. Structured output for downstream systems.
All three engines are installed as a shared Python package. Applications import them. Domain data lives in the application, not in the engine. This means the same codebase processes engineering specs, Bible passages, and marketing corpora without modification.
The foundation
Every engine depends on Decompose, an open-source deterministic text classifier published on PyPI as decompose-mcp. It classifies any text into structured semantic units (authority, risk, attention, entities) without an LLM. Same input, same output, every time. 13.9ms average. 63 passing tests. Zero runtime dependencies.
Decompose is important enough to give away. It proves the architectural principle (deterministic before probabilistic) and serves as the entry point for developers who want to understand how the platform works.
Trained models
Three custom models, fine-tuned with MLX LoRA on local hardware. No cloud training. No data leaves the machine.
| Model | Training Pairs | Status |
|---|---|---|
| echology:latest | 3,081 | Deployed. Platform identity + curated supplemental data. |
| aecai:latest | 1,173 | Data synced. Retrain queued after echology completes. |
| open-scripture:latest | 41,684 | Deployed. Val loss 0.381. 66 books, 9,971 passages. |
The three-model architecture prevents cross-contamination. The platform model knows echology. The AEC model knows engineering. The scripture model knows the Bible. None of them hallucinate expertise they weren't trained on.
Applications
Five applications built on the platform today:
- AECai: Document intelligence for architecture, engineering, and construction firms. FastAPI server, 10 route modules, 6 Temporal workflows, 569 tests. Deployed on local hardware via launchd.
- Geode: AI visibility optimization. Makes organizational expertise citable by ChatGPT, Perplexity, Google AI Overviews, and Copilot. Citation baseline measured: tracking 16 queries across 5 platforms for 3 brands.
- Open Scripture: AI-ready Bible dataset. 31,100 verses, 344,799 cross-reference edges, 8 theological disciplines. The same pipeline that processes engineering specs processes scripture with zero code changes.
- RBS Demo: Insurance policy QC. Compares documents across sources and flags discrepancies. Three-tier extraction: regex first, decompose + LLM second, vision model third.
- Ops: Internal command center. 53 dashboard routes, citation tracking, lead pipeline, training data management, corpus ingestion. SQLite backend. Not public.
Infrastructure
Everything runs on two machines. A Mac Mini (Apple Silicon) handles the operational backend: Ollama with 6 models, Qdrant vector database, the ops dashboard, and the aecai server via a Cloudflare tunnel. A Surface Laptop handles development. Public-facing sites (echology.io, aecai.io, geode.digital) are static HTML on GitHub Pages, so they never depend on a single machine being online.
630+ tests pass across the ecosystem. Zero cloud dependencies. All data processing happens on hardware the user controls.
What's next
We are in Phase 1: proving the platform works end-to-end. The engines are built. The models are trained. The applications are running. The next step is demonstrating measurable outcomes.
- Geode citation re-audit (March 21), first measurement of whether structured content changes AI citation behavior. Baseline: Geode 0/6, echology 1/5, AECai 2/5.
- AECai model retrain, platform base + AEC domain data, expanded with the improved echology:latest foundation.
- First customer pipeline test, real geotechnical engineering project data processed through the full customer deployment workflow.
Phase 2 is applying proven results to the AECai vertical. Phase 3 is marketing automation powered by the ops layer. Each phase validates the previous one. Nothing ships until it's demonstrated.
Principle-based architecture compounds. Feature-based architecture plateaus. We are building the kind that compounds.