The anatomy of an autonomous system.
Read this as an operating manual. Every layer, agent and loop exists to let intelligence discover, validate and improve — without a human in the inner loop, and never outside of governance.
We do not build a program. We grow a system.
It is easier to reason about the whole as a living thing — with a body, a brain, a nervous system, senses, an immune system and memory. Select an organ.
Body
→InfrastructureThe substrate everything runs on.
Distributed compute, low-latency data fabric, storage and orchestration. The physical form that gives the intelligence somewhere to live, remember and act.
Five disciplines, one intelligence.
No single technique is sufficient. Each contributes a distinct capability; the Model Context Protocol lets them compose.
Machine Learning
Statistical learning to surface durable relationships within high-dimensional, low signal-to-noise financial data.
Deep Learning
Neural architectures that learn representations of sequence, structure and regime that classical features cannot express.
Reinforcement Learning
Agents that learn policies under uncertainty — sizing, timing and execution as decisions made under consequence, not prediction alone.
Agentic AI
Goal-directed agents that plan, use tools and coordinate to run the research loop with minimal human instruction.
Model Context Protocol
The shared language of context and capability that lets models, tools and data compose into a single, governed system.
From compute to self-learning.
Each layer depends on the one below and improves the ones around it. Expand a layer to read its role.
Compute, data fabric and orchestration.
The foundation: distributed compute, streaming data, storage and the scheduling that keeps thousands of experiments in flight.
Eight specialists. One research loop.
Autonomy is not one model doing everything. It is a coordinated network of specialised agents, each accountable for one part of the scientific method — bound together by the Model Context Protocol.
Hypothesis Agent
Proposes testable ideas from data, theory and prior failures.
Data Agent
Sources, cleans and represents the world the system reasons over.
Feature Agent
Constructs and prunes the representations models learn from.
Validation Agent
Adjudicates ideas with out-of-sample and adversarial testing.
Portfolio Agent
Composes validated signals into portfolios under constraints.
Execution Agent
Turns intent into orders while minimising impact and cost.
Risk Agent
Enforces limits, audits behaviour and holds veto authority.
Meta-Learning Agent
Studies the system itself and directs where it should improve.
A loop that never resets.
Every idea travels the same path — and every outcome feeds the next. This is how understanding compounds.
- 1
Hypothesise
An idea is proposed with an explicit, falsifiable claim.
- 2
Test
The claim is run against unseen data and hostile scenarios.
- 3
Validate
Statistics — not intuition — decide what survives.
- 4
Deploy
Survivors become governed, versioned live signals.
- 5
Observe
Behaviour is monitored for drift, decay and regime shift.
- 6
Learn
Every outcome updates priors and seeds the next hypothesis.
Long-term vision
A research organism that improves faster than the edge it discovers decays — designed to outlast every model inside it.
Components will be replaced. Techniques will be superseded. The architecture — how ideas are formed, judged, governed and remembered — is what we intend to endure.