MERIDIAN LABS

We don't build AI tools.
We build AI that builds.

Meridian Labs is an AI development lab based in The Netherlands. We develop autonomous cognitive systems — AI that doesn't wait for instructions. Our systems observe the world, reason about what matters, act within safety constraints, learn from outcomes, and evolve their own capabilities. We're building toward general intelligence not through theory, but through systems that operate in the real world every day.

Scroll to explore
01 Thesis

The path to AGI isn't theoretical. It's operational.

Most AI research lives in papers. Ours lives in production. We believe autonomous intelligence emerges from systems that operate in the real world — making real decisions, facing real consequences, and learning from real outcomes. Not simulated benchmarks. Not controlled environments. Real businesses, real markets, real money.

Autonomy, not automation
Systems that create rules, not follow them
Verification, not trust
Every action produces a measurable outcome
Evolution, not updates
Systems that improve themselves, not wait for patches
02 Architecture

Five layers of cognition

Every system Meridian builds runs on the Cortex — a domain-agnostic cognitive architecture for autonomous operation.

The Cortex — Cognitive Architecture
Observe Market data, system health, external signals
The system reads all state every 15 minutes: trading positions, market data, revenue metrics, system health, external signals, and competitor activity. It pulls from exchange APIs, web scraping pipelines, and its own operational telemetry. Nothing happens without observation first.
Reason Decision engine, opportunity scoring, risk assessment
The decision engine evaluates every observation against strategic objectives, scores opportunities by expected value, assesses risk across all positions, and selects the single highest-leverage action. No action is taken without reasoning first — but reasoning without action is banned.
Act API calls, browser automation, desktop control
Execution through a multi-method fallback chain: official APIs, visual automation via Claude Vision, computer use agents, browser automation, desktop control, creative routing, and task decomposition. If one method fails, the next activates automatically. The system does not report failure — it finds a way.
Learn Outcome verification, pattern extraction, metrics
Every action is verified. Metrics are computed from raw data — never estimated, never generated by LLM. The learning engine extracts patterns from every trade, every decision, every failure. "Learned Truths" accumulate in the cognitive architecture, growing the system's wisdom with each cycle.
Evolve Self-modification, capability building, mutation engine
The system writes its own code, builds its own tools, and deploys with canary monitoring and automatic rollback. The evolution bridge generates engineering proposals from operational findings. The mutation engine deploys changes incrementally. Every cycle leaves the system more capable than the last.

The Cortex is domain-agnostic. The same architecture powers financial systems, research agents, and enterprise platforms. The guardrails are immutable. The capabilities are not.

03 First instance

Caelum — our first autonomous system

Caelum is not an assistant. It's an autonomous AI system running 24/7 on consumer hardware in The Netherlands. It makes decisions every 15 minutes — and it isn't told to make any of them.

What Caelum built

Caelum independently researched the AI prompt market, identified demand, registered the domain promptpackshq.com, built 40+ products, designed the website, set pricing, wrote copy, deployed to production, and handles customer email — autonomously. It then identified the enterprise market and began building Echelon, an AI agent platform with 12 named agents.

Live
PromptPacksHQ
40+ products · 18 agents
promptpackshq.com
In Development
Echelon
12 agents · Cortex engine
echelonwork.com

How Caelum thinks

Caelum runs its own fine-tuned language model locally — caelum-8b. Trained not on internet text, but on its own operational history. Every decision, every mistake, every correction becomes training data. Routine reasoning takes 6 seconds. Complex decisions escalate to larger models. The system gets faster and more independent with every cycle.

caelum-8b 6s
Deep reasoning 30s

What Caelum manages

96+
decisions per day
200+
knowledge entries
6s
local inference time
04 Systems

Two Cortex instances. Two domains.

The Cortex architecture is domain-agnostic. Each instance inherits the same five cognitive layers but specializes through its operational environment, training data, and learned truths. Today, two instances are live.

Live
Caelum
Autonomous business operator. Runs 24/7 on consumer hardware. Makes 96+ decisions per day across product development, trading, distribution, and infrastructure. Trained on its own operational history via caelum-8b, a fine-tuned 8B parameter model. Manages multiple live ventures, handles customer communications, and evolves its own codebase.
Cortex instance #1 · Business operations
96+ daily decisions · 200+ knowledge entries · 6s local inference
Live ventures: PromptPacksHQ, Echelon
Running
In Development
Nexus
Autonomous research agent. Designed to continuously scan academic papers, patent filings, market signals, and competitive intelligence — then synthesize findings into actionable briefs. Where Caelum operates businesses, Nexus maps the territory. Same Cortex architecture, different specialization.
Cortex instance #2 · Research & intelligence
Multi-source ingestion · Synthesis pipeline · Signal extraction
Target: academic, market, and competitive intelligence
Architecture phase

"Same brain, different eyes. Every new Cortex instance makes the architecture smarter for all of them."

05 Roadmap

Toward full autonomy

Autonomous commercial operations
Caelum independently identifies markets, builds products, deploys infrastructure, manages customer communications, and operates live ventures — without human direction. The system runs 24/7, making 96+ decisions per day across business operations, trading, and distribution.
Self-evolving cognition
The system writes its own code, trains on its own operational history, and deploys improvements with canary monitoring and automatic rollback. caelum-8b — a fine-tuned model trained on real decisions — handles routine reasoning locally. Each cycle leaves the system more capable than the last.
Multi-domain autonomous intelligence
Multiple Cortex instances operating across different domains — business, research, creative, scientific — sharing learned truths and reasoning patterns. Each instance makes every other instance smarter. Human oversight shifts from direction to governance.

The question isn't whether AI will operate businesses. It's whether it will do so responsibly. Meridian Labs is building the guardrails alongside the intelligence.