/ Sovereign operations field

Not a chatbot layer. An operating surface.

CASTLE is designed as a room-scale system: identity, policy, memory, model routing, and evidence all move together. The site now shows that shape directly instead of asking visitors to imagine it from copy.

/ 01

Operator request enters the room.Shield verifies identity and Tower applies policy before model work begins.

/ 02

Context resolves locally.Crown and Atlas retrieve institutional memory without leaving customer infrastructure.

/ 03

Every decision leaves evidence.Knox records the route, source set, and completion lineage for review.

Operator

Signed answer

Why this market is ours

The choices frontier labs already made.

Each row below is a strategic decision OpenAI, Anthropic, and Google have already committed to — choices that compound their margins and exclude them from the institutions whose work matters most. The same decisions are exactly why CASTLE exists.

/ 01 — GeneralityTheir product must serve everyone.

The frontier-lab business model is paid by the token across millions of users. Specializing for a single regulated vertical shrinks TAM and breaks unit economics. CASTLE is purpose-built for institutional cognitive work — every parameter earns its place on Canadian, regulated, audit-graded work.

/ 02 — Hosting modelInference must run on their GPUs.

Frontier-lab margins depend on serving inference from their own datacenters. Deploying on customer iron breaks that model. CASTLE runs entirely on customer infrastructure — on-prem, private cloud, or air-gapped — because we charge for the room, not the rent on the room.

/ 03 — Data postureThey cannot ingest your data.

Frontier labs legally and commercially cannot hold customer-confidential corpora for fine-tuning at the scale a real institution requires. CASTLE compounds on your data, in your environment, under your policy — a capability frontier labs structurally cannot offer.

/ 04 — ProvenanceThey don't track inference-level lineage.

Consumer users don't demand provenance — so frontier labs never built it. Knox signs every hop, every header, every hash, append-only. Reconstructable to any reviewer with the right key. Required by every regulator we'll ever meet.

/ 05 — CompletenessTheir stack stops at the API.

Selling the full operating stack — model, memory, data, identity, policy, audit — would compete with their own infrastructure-startup customers. CASTLE is the complete stack as one platform: customers stop integrating six vendors and start operating one system.

/ 06 — Institutional memoryTheir privacy posture forbids retention.

The frontier-lab privacy story is "we don't keep your interactions." That's the right answer for consumers — and exactly the wrong answer for an organization whose memory IS its competitive advantage. CASTLE remembers across every session, every operator, every quarter, on your hardware.

Capability gets us in the door. Completeness keeps us there. Sovereignty is what makes us impossible to displace once deployed.

/ First principles

We didn't ship faster.
We built from the ground up.

Every architectural decision in AXE was made from first principles — not from what the rest of the AI industry shipped that quarter. Where the orthodox answer was wrong for institutions, we built the right answer instead. Six of those decisions, restated below.

/ 01 — The model

We didn't fine-tune. We trained.

Every Edge weight was learned from scratch on hardware in Canada. A model built on someone else's foundation inherits someone else's posture toward your data. Provenance starts at random initialization or it doesn't start at all.

/ 02 — The memory

We chose ontology, not RAG.

Vector retrieval is a parlor trick for unstructured search. Institutional work runs on relationships — Campaign → Audience → Channel → Performance. Nouns without verbs cannot drive operations. Crown ships verbs from day one.

/ 03 — The substrate

We chose Postgres over Cassandra.

Hyperscaler reference architectures default to systems built for scale we don't have. The simplest correct system at our scale is faster than the most scalable wrong one. Atlas is Postgres + pgvector + Row-Level Security. We earn the right to add complexity.

/ 04 — The audit chain

Append-only, post-quantum from boot.

Frontier labs don't track inference-level provenance because consumers don't demand it. Every regulator we'll meet does. Knox signs with Ed25519 and ML-DSA-65 — the day quantum breaks Ed25519 isn't a day we plan to scramble.

/ 05 — The trust model

Zero-trust from boot, not from launch.

"Trust then verify" is how supply-chain worms win. Shield verifies every actor, every device, every scope on every request — not because policy demands it, but because the cheapest place to stop an attacker is before the first byte they want to read.

/ 06 — The deployment

On your iron, or not at all.

Customer hardware. Customer network. Customer key material from boot. Air-gap capable by default. The frontier-lab business model requires inference on their GPUs — which is why customers whose work doesn't fit their TOS structurally cannot use them. Different room. Different rules.

The right answer is rarely the cheapest one. We are betting that institutions whose work matters will eventually demand both.

— James Lewis · CEO · Built in Canada, for the rooms that can't use frontier APIs.

See Echo in Action

Echo Terminal — Canadian Inference

$ curl https://echo.castle.local/v1/chat/completions \

-H "X-AXE-Key: axe_mk_••••••••" \

-d '{"model": "edge-1", "stream": true,

"messages": [{"role": "user", "content": "Analyze Q3 defense procurement trends"}]}'

{"choices":[{"delta":{"content":"Based on the latest PSPC procurement data, "

{"choices":[{"delta":{"content":"three key trends emerge in Q3 2026: "

{"choices":[{"delta":{"content":"1) Increased emphasis on Canadian-built solutions "

{"choices":[{"delta":{"content":"following the Buy Canadian Policy... "

✓ 847 tokens · edge-1 · 1.2s · logged to training pipeline

Our Software

The Canadian inference engine. Routes queries to the right model, executes tools, streams responses. OpenAI-compatible.

/0.1 — E

Echo

Institutional memory and knowledge vault. Vector search, graph-based knowledge, every interaction captured and retrievable.

/0.2 — C

Crown

Canadian database infrastructure. PostgreSQL with row-level security, object storage, real-time subscriptions. Every byte stays on your hardware.

/0.3 — A

Atlas

Zero-trust authentication across the entire platform. Every request verified. No credentials, no access.

/0.4 — S

Shield

Granular permission control. Role-based and attribute-based access down to individual records and endpoints.

/0.5 — T

Tower

AI-controlled Chromium browser. Navigate, extract, and interact with the web under strict supervision. Every action logged.

/0.6 — L

Lens

The Request Lifecycle

Every request authenticated. Every action logged. Every interaction improving the system.

User

Shield

Tower

Echo

Crown

Atlas

/ Architecture · How CASTLE Works

Four diagrams.
The whole platform.

No black boxes. No secrets. The shape of what you'd be deploying — drawn so an operator, a CFO, and a minister can each see what they need to see.

/ 01 — Platform

The Constellation

One model at the centre. Six systems around it. One audit chain wrapping the whole thing. Every request crosses the audit boundary in both directions — nothing enters or leaves CASTLE without a receipt.

EDGE · 6 SYSTEMS · KNOX

/ 02 — Model

Edge: one family.

Every weight learned from scratch on Canadian iron. One training pipeline. One lineage. Four sizes — pick the one your iron can run, keep the option to grow.

Not fine-tuned from someone else's model. Not API-wrapped. Trained.

FROM SCRATCH · ON-PREM · PORTABLE

/ 03 — Algorithm

Intent. Memory. Action.

The AXE Algorithm reads what you actually meant, asks memory what's already known, decides what to do, then proves what it did. Inference is the cheap part.

The interesting layer is the routing — and the receipt that follows it.

DECISIONS · NOT JUST TOKENS

/ 04 — Memory

The moat is memory.

Inference forgets. Organizations don't. CASTLE's memory has four layers — from the rules that never change to the session that ended five minutes ago — and every layer is searchable, auditable, yours.

Strip memory from an AI and you're left with autocomplete. We don't ship autocomplete.

4 LAYERS · ON YOUR HARDWARE

Traditional Cloud AI

Data Custody

Your data on their servers. Their rules.

With CASTLE

Data Custody

100%

Your hardware. Your models. Your rules. Zero external dependencies.

Impact

Enterprise Deployment

IMI International

First production deployment of CASTLE underway. Six integrated systems, one Canadian-trained model family, zero cloud dependency, complete data custody.

6 Components · 100% Canadian

Edge Model Family

Trained on Canadian Iron

Every weight learned from scratch on hardware in Canada. One model family, four sizes — from laptop to cluster. No upstream dependency.

From Scratch · One Family · Yours

Sprint Velocity

CASTLE Day One

32 tasks completed in a single sprint day. Autonomous team execution across infrastructure, security, inference, and documentation.

32 Tasks · 63% Complete · Day 1

01 —

Echo

The Canadian Inference Engine

Explore →

02 —

Crown

Institutional Memory & Knowledge

Explore →

03 —

Atlas

Canadian Data Infrastructure

Explore →

04 —

Shield

Zero-Trust Authentication

Explore →

Compliance & Certifications

FIPS 140-2 Path

Protected B Ready

PIPEDA Compliant

Canadian Built

Air-Gap Capable

What Makes CASTLE Powerful

Day 1 Custody

Deploy on your hardware in under 4 hours. Full control from day one.

AI Without Dependency

Your models, your data, zero cloud calls. Complete independence.

Full Stack Integration

Six systems, one unified API. Inference, memory, auth, storage, browsing, and orchestration.

Canadian Built

FIPS 140-2 path, Protected B ready. Your data stays in Canada.

What Leaders Say

If you're not invited to the table, you're probably on the menu. We built CASTLE so Canadian organizations never have to depend on someone else's infrastructure for their intelligence layer.

James Lewis

CEO & Founder, AXE Technologies

Custody isn't a feature you can toggle on. It's the architecture. CASTLE was designed from first principles to eliminate every external dependency.

Michael Mayo

COO & Co-Founder, AXE Technologies

For the first time, we have AI capability that meets our data custody requirements without compromising on performance or functionality.

(Anonymous)

Senior Technology Officer, Canadian Federal Agency

Software that runs intelligence on hardware you own.

The operating system institutions actually need.

Protected B-ready. FIPS-track. Canadian-soil.

Inference where your data already lives.

Air-gapped capable. Zero external trust.

Not a chatbot layer. An operating surface.

The choices frontier labs already made.

We didn't ship faster.
We built from the ground up.

We didn't fine-tune. We trained.

We chose ontology, not RAG.

We chose Postgres over Cassandra.

Append-only, post-quantum from boot.

Zero-trust from boot, not from launch.

On your iron, or not at all.

See Echo in Action

Our Software

The Request Lifecycle

Four diagrams.
The whole platform.

The Constellation

Edge: one family.

Intent. Memory. Action.

The moat is memory.

Traditional Cloud AI

With CASTLE

Impact

What Makes CASTLE Powerful

What Leaders Say

Interested in deploying Canadian AI
for your organization?

Talk to .

Software that runs intelligence on hardware you own.

The operating system institutions actually need.

Protected B-ready. FIPS-track. Canadian-soil.

Inference where your data already lives.

Air-gapped capable. Zero external trust.

Not a chatbot layer. An operating surface.

The choices frontier labs already made.

We didn't ship faster.We built from the ground up.

We didn't fine-tune. We trained.

We chose ontology, not RAG.

We chose Postgres over Cassandra.

Append-only, post-quantum from boot.

Zero-trust from boot, not from launch.

On your iron, or not at all.

See Echo in Action

Our Software

The Request Lifecycle

Four diagrams.The whole platform.

The Constellation

Edge: one family.

Intent. Memory. Action.

The moat is memory.

Traditional Cloud AI

With CASTLE

Impact

What Makes CASTLE Powerful

What Leaders Say

Interested in deploying Canadian AIfor your organization?

Talk to .

We didn't ship faster.
We built from the ground up.

Four diagrams.
The whole platform.

Interested in deploying Canadian AI
for your organization?