ENTERPRISE INFRASTRUCTURE
Otonomii's Brain
Otonomii's Brain
Otonomii builds perpetual machines: autonomous systems that discover, learn and adapt in complex market environments.
We are a machine learning research organization that happens to operate in markets because markets provide the richest, most adversarial learning environment available.
Otonomii builds perpetual machines: autonomous systems that discover, learn and adapt in complex market environments.
We are a machine learning research organization that happens to operate in markets because markets provide the richest, most adversarial learning environment available.
Platform
Platform Overview
Otonomii is an enterprise AI platform built on the principle that complexity is the enemy of durability. Where most enterprise
AI platforms add features to chase capability, Otonomii removes them to achieve resilience.
The platform is designed around a single observation: the human brain is the most complex machine in existence, yet it
evolved not to maximize complexity but to reduce it and to generalize, compress and optimize energy expenditure.
Otonomii provides four cognitive layers: Brain, Mind, Machine and Arena, that snap together to create autonomous
systems capable of perceiving, understanding, deciding and acting.
The platform is domain agnostic by design. The same cognitive architecture that operates in financial markets can optimize supply chains, manage security operations or orchestrate complex workflows.
Architecture Deep-Dive
The four layer architecture enforces strict dependency rules. Each layer can only import from layers below it, while two layers, Brain and Arena, import nothing at all. This constraint prevents circular dependencies, ensures testability in isolation and makes the system's behavior predictable even as it grows in capability.
The four layer architecture enforces strict dependency rules. Each layer can only import from layers below it, while two layers, Brain and Arena, import nothing at all. This constraint prevents circular dependencies, ensures testability in isolation and makes the system's behavior predictable even as it grows in capability.
Brain
(imports nothing)
Domain Agnostic Memory Substrate
The Brain layer is the foundational memory substrate upon which all higher cognition is built. It is deliberately domain agnostic and it knows nothing about markets, language or any specific problem domain. Its sole purpose is to store, retrieve, associate and compress patterns. Write commits new observations to memory. Lookup retrieves relevant prior experience given a query. Link creates associative connections between related memories, building a web of
relationships that mirrors how biological memory works. Consolidate periodically compresses and reorganizes stored patterns, pruning redundancy while preserving the essential structure. This layer is inspired by the hippocampal memory system: a general-purpose recording mechanism that captures experience without interpretation.
The Brain layer is the foundational memory substrate upon which all higher cognition is built. It is deliberately domain agnostic and it knows nothing about markets, language or any specific problem domain. Its sole purpose is to store, retrieve, associate and compress patterns. Write commits new observations to memory. Lookup retrieves relevant prior experience given a query. Link creates associative connections between related memories, building a web of
relationships that mirrors how biological memory works. Consolidate periodically compresses and reorganizes stored patterns, pruning redundancy while preserving the essential structure. This layer is inspired by the hippocampal memory system: a general-purpose recording mechanism that captures experience without interpretation.
Mind
(imports Brain only)
Semantic Understanding Through Prediction Error
The Mind layer transforms raw memory into understanding. Encode converts raw observations into internal representations. Predict generates expectations about what should happen next based on past patterns. Error measures the deviation between prediction and reality, This prediction error is the most important learning input in the entire system. Learn updates internal models based on prediction errors, strengthening accurate patterns and weakening inaccurate ones. Concepts emerge naturally as clusters of related patterns that the system discovers through repeated exposure. Votes allow multiple internal models to express opinions about the current state, creating an ensemble of perspectives. Salience determines which signals deserve attention and which can be safely ignored. The
Mind layer embodies a core insight from constructed emotion theory: the brain is fundamentally a prediction machine and surprise drives all learning.
The Mind layer transforms raw memory into understanding. Encode converts raw observations into internal representations. Predict generates expectations about what should happen next based on past patterns. Error measures the deviation between prediction and reality, This prediction error is the most important learning input in the entire system. Learn updates internal models based on prediction errors, strengthening accurate patterns and weakening inaccurate ones. Concepts emerge naturally as clusters of related patterns that the system discovers through repeated exposure. Votes allow multiple internal models to express opinions about the current state, creating an ensemble of perspectives. Salience determines which signals deserve attention and which can be safely ignored. The
Mind layer embodies a core insight from constructed emotion theory: the brain is fundamentally a prediction machine and surprise drives all learning.
Machine
(imports Brain + Mind)
The Agent That Decides and Acts
The Machine layer is where cognition becomes behavior. Threshold determines the confidence level required before acting too low and the system acts on noise, too high and it misses opportunities. The threshold is not fixed; it adapts based on regime, risk context and recent performance. Stopping Rule defines when to cease an activity and
when to close a position, when to stop searching for more information, when to accept the current answer as good enough. This is perhaps the most underappreciated capability in autonomous systems. Action translates decisions into concrete operations in the external world. Resources manages computational budgets, API quotas, memory
allocation and time constraints. The Machine layer enforces the principle that every action must pass through both a confidence gate and a risk assessment before execution.
The Machine layer is where cognition becomes behavior. Threshold determines the confidence level required before acting too low and the system acts on noise, too high and it misses opportunities. The threshold is not fixed; it adapts based on regime, risk context and recent performance. Stopping Rule defines when to cease an activity and
when to close a position, when to stop searching for more information, when to accept the current answer as good enough. This is perhaps the most underappreciated capability in autonomous systems. Action translates decisions into concrete operations in the external world. Resources manages computational budgets, API quotas, memory
allocation and time constraints. The Machine layer enforces the principle that every action must pass through both a confidence gate and a risk assessment before execution.
Arena
(imports nothing)
World Model, Independent of Machines
The Arena layer represents the external world that machines operate in. It is deliberately independent of the Machine layer and the world exists regardless of who or what is acting in it. World Model maintains a structured representation of the current state of the environment, including regime classification, market microstructure and contextual factors.
Charter defines the rules of engagement such as what is permitted, what is forbidden, risk limits and operating boundaries. Instruments are the specific tools available for interaction with the world's financial instruments, data feeds and execution venues. By separating the Arena from the Machine, we ensure that our understanding of the world is not contaminated by our desire to act in it. Multiple machines can share the same Arena, each with different strategies but a common understanding of reality.
The Arena layer represents the external world that machines operate in. It is deliberately independent of the Machine layer and the world exists regardless of who or what is acting in it. World Model maintains a structured representation of the current state of the environment, including regime classification, market microstructure and contextual factors.
Charter defines the rules of engagement such as what is permitted, what is forbidden, risk limits and operating boundaries. Instruments are the specific tools available for interaction with the world's financial instruments, data feeds and execution venues. By separating the Arena from the Machine, we ensure that our understanding of the world is not contaminated by our desire to act in it. Multiple machines can share the same Arena, each with different strategies but a common understanding of reality.
Import Rules
Brain
Imports nothing
Mind
Brain only
Machine
Brain + Mind
Arena
Imports nothing
Import rules are enforced by an AST based scanner that runs on every commit. Violations are blocked at the CI level. Data is a cross cutting concern and imports nothing from any layer.
Import rules are enforced by an AST based scanner that runs on every commit. Violations are blocked at the CI level. Data is a cross cutting concern and imports nothing from any layer.
Hub and Spoke Architecture
The Hub and Spoke architecture is a convergence pattern inspired by how the brain integrates information from multiple sensory modalities. Independent sensory spokes process data from different domains and each operating autonomously with its own memory, normalization and recall mechanisms. The convergence hub receives assessments from all spokes and performs a single critical function: detecting agreement, conflict and novelty across them.
The Hub and Spoke architecture is a convergence pattern inspired by how the brain integrates information from multiple sensory modalities. Independent sensory spokes process data from different domains and each operating autonomously with its own memory, normalization and recall mechanisms. The convergence hub receives assessments from all spokes and performs a single critical function: detecting agreement, conflict and novelty across them.
The hub does not care what the spokes process. It does not need to understand the internal logic of any individual spoke. It only asks three questions: Do the spokes agree? If so, that is a increase confidence. Do the spokes conflict? If so, that is a red flag and critically, conflict between spokes is the information. It means something unusual is happening that warrants attention. Do some spokes have no opinion? That is a caution indicator and insufficient information to form a view.
The hub does not care what the spokes process. It does not need to understand the internal logic of any individual spoke. It only asks three questions: Do the spokes agree? If so, that is a increase confidence. Do the spokes conflict? If so, that is a red flag and critically, conflict between spokes is the information. It means something unusual is happening that warrants attention. Do some spokes have no opinion? That is a caution indicator and insufficient information to form a view.
This architecture is deliberately non-hierarchical. No spoke is more important than another by default. Importance is determined dynamically through salience weighting, which adjusts based on the current regime and the historical accuracy of each spoke in similar conditions.
This architecture is deliberately non-hierarchical. No spoke is more important than another by default. Importance is determined dynamically through salience weighting, which adjusts based on the current regime and the historical accuracy of each spoke in similar conditions.
Spoke Processing Pipeline
01 Input
Raw data enters the spoke from its designated source such as price feeds, news, sentiment, volume or any other sensory channel.
Raw data enters the spoke from its designated source such as price feeds, news, sentiment, volume or any other sensory channel.
02 Normalize
Each spoke converts its input into a standardized internal representation. This normalization allows the hub to compare signals
from fundamentally different domains.
Each spoke converts its input into a standardized internal representation. This normalization allows the hub to compare signals
from fundamentally different domains.
03 Store
The spoke maintains its own memory of past observations, building a history specific to its domain. This local memory enables
pattern recognition within the spoke.
The spoke maintains its own memory of past observations, building a history specific to its domain. This local memory enables
pattern recognition within the spoke.
04 Recall
When queried, the spoke retrieves relevant historical patterns and generates its independent assessment of the current situation.
When queried, the spoke retrieves relevant historical patterns and generates its independent assessment of the current situation.
Hub Convergence Signals
Agreement
Multiple spokes arrive at the same conclusion independently. Confidence increases proportionally with the number of agreeing spokes and their
historical reliability.
Multiple spokes arrive at the same conclusion independently. Confidence increases proportionally with the number of agreeing spokes and their
historical reliability.
Conflict
Spokes disagree on the current state or expected outcome. This is not an error, it is the most valuable signal. Conflict indicates that conditions
are unusual or that a regime transition may be underway.
Spokes disagree on the current state or expected outcome. This is not an error, it is the most valuable signal. Conflict indicates that conditions
are unusual or that a regime transition may be underway.
No Opinion
One or more spokes cannot form a view. This silence is informative, It means the current situation falls outside the training distribution of that
spoke. Caution is warranted.
One or more spokes cannot form a view. This silence is informative, It means the current situation falls outside the training distribution of that
spoke. Caution is warranted.
Hierarchical Pattern-Outcome Trees
Patterns do not exist in isolation. Every pattern resolves into an outcome and every outcome is the next pattern. This creates a tree structure that never stops branching, an ever expanding web of conditional relationships that captures the sequential, path dependent nature of complex systems.
Patterns do not exist in isolation. Every pattern resolves into an outcome and every outcome is the next pattern. This creates a tree structure that never stops branching, an ever expanding web of conditional relationships that captures the sequential, path dependent nature of complex systems.
The critical insight is that the meaning of the current pattern depends on which branch of the tree you arrived from. A breakout above resistance has different implications depending on whether it followed a long consolidation, a sharp reversal or a gap open. The same surface level pattern can mean entirely different things depending on the path that led to it. Context is not supplementary, It is constitutive.
The critical insight is that the meaning of the current pattern depends on which branch of the tree you arrived from. A breakout above resistance has different implications depending on whether it followed a long consolidation, a sharp reversal or a gap open. The same surface level pattern can mean entirely different things depending on the path that led to it. Context is not supplementary, It is constitutive.
Prediction error is the most important signal in this framework. When the tree branches in a direction that was not predicted, that deviation carries maximum information. It signals that the model's understanding of the causal structure is incomplete and needs updating.
Prediction error is the most important signal in this framework. When the tree branches in a direction that was not predicted, that deviation carries maximum information. It signals that the model's understanding of the causal structure is incomplete and needs updating.
This tree structure also enables a form of hierarchical planning. By projecting forward along expected branches, the system can reason about likely future states and prepare contingencies for alternative branches.
This tree structure also enables a form of hierarchical planning. By projecting forward along expected branches, the system can reason about likely future states and prepare contingencies for alternative branches.
Deployment
Options
Deployment Options
Cloud Deployment
AWS, GCP, Azure
Fully managed deployment on major cloud platforms. Auto scaling inference clusters, managed data pipelines and integrated monitoring. Regional deployment options ensure data residency compliance. Multi region failover for mission critical workloads.
Fully managed deployment on major cloud platforms. Auto scaling inference clusters, managed data pipelines and integrated monitoring. Regional deployment options ensure data residency compliance. Multi region failover for mission critical workloads.
On-Premise
Your Infrastructure
Complete platform deployment within your data center. Air gapped options available for regulated industries. Full control over hardware, networking and security perimeter. No external dependencies once deployed.
Complete platform deployment within your data center. Air gapped options available for regulated industries. Full control over hardware, networking and security perimeter. No external dependencies once deployed.
Hybrid
Mixed Infrastructure
Training and experimentation in the cloud, inference and decision making on premise. Sensitive data never leaves your perimeter. Cloud resources handle burst computational needs. Seamless synchronization between environments.
Training and experimentation in the cloud, inference and decision making on premise. Sensitive data never leaves your perimeter. Cloud resources handle burst computational needs. Seamless synchronization between environments.
Data
Residency
and Regional
Inference
Data Residency and Regional Inference
All deployment options support configurable data residency controls. Data can be pinned to specific geographic regions, ensuring compliance with GDPR, data sovereignty requirements and industry specific regulations.
All deployment options support configurable data residency controls. Data can be pinned to specific geographic regions, ensuring compliance with GDPR, data sovereignty requirements and industry specific regulations.
Inference can be performed regionally in order to be close to the data source and reducing latency and eliminating cross border data transfers. For multi region deployments, the platform supports active configurations where each region operates independently while maintaining eventual consistency for non critical metadata.
Inference can be performed regionally in order to be close to the data source and reducing latency and eliminating cross border data transfers. For multi region deployments, the platform supports active configurations where each region operates independently while maintaining eventual consistency for non critical metadata.
Integration
REST APIs
Full CRUD operations on all platform resources. OpenAPI 3.1 specifications. Rate limiting, pagination and versioning built in.
Full CRUD operations on all platform resources. OpenAPI 3.1 specifications. Rate limiting, pagination and versioning built in.
SDKs
Native libraries for Python, TypeScript, Go and Java. Type safe interfaces that mirror the platform’s internal architecture.
Native libraries for Python, TypeScript, Go and Java. Type safe interfaces that mirror the platform’s internal architecture.
Webhooks
Real-time event notifications for state changes, predictions, alerts and anomalies. Configurable filters and retry logic.
Real-time event notifications for state changes, predictions, alerts and anomalies. Configurable filters and retry logic.
Custom Connectors
Build integrations to proprietary systems using the connector SDK. Schema-driven data mapping with automatic validation.
Build integrations to proprietary systems using the connector SDK. Schema-driven data mapping with automatic validation.
Multi-Model Orchestration
Route tasks to the right model based on task type: reasoning, consensus, algorithms, long-running execution. Cost-optimized by default.
Route tasks to the right model based on task type: reasoning, consensus, algorithms, long-running execution. Cost-optimized by default.
Streaming
Server-sent events and WebSocket interfaces for real-time data flows. Backpressure handling and automatic reconnection.
Server-sent events and WebSocket interfaces for real-time data flows. Backpressure handling and automatic reconnection.
Strategy Survival
Analysis
Strategy Survival Analysis
1,878
Total Strategies Tested
990
Experiment 1 Strategies
0
Exp 1 Survivors
888
Experiment 2 Strategies
12
Exp 2 Survivors
1.35%
Survival Rate
Experiment 1: Regime-Blind
990 strategies tested using standard pivot points without regime awareness. Zero survivors. When volatility collapsed by 3x mid experiment, every strategy failed.
990 strategies tested using standard pivot points without regime awareness. Zero survivors. When volatility collapsed by 3x mid experiment, every strategy failed.
The strategies could not distinguish between the regime they were designed for and the regime they were operating in. This confirmed that prediction accuracy without regime context is meaningless.
The strategies could not distinguish between the regime they were designed for and the regime they were operating in. This confirmed that prediction accuracy without regime context is meaningless.
Experiment 2: Regime-Aware
888 strategies tested using money zones, including VPOC, Fibonacci, CPR and multi timeframe pivots, with full regime awareness. 12 survivors with a 1.35% survival rate.
888 strategies tested using money zones, including VPOC, Fibonacci, CPR and multi timeframe pivots, with full regime awareness. 12 survivors with a 1.35% survival rate.
The survivors shared a common trait: they were highly selective, only engaging when regime conditions matched their design parameters. Selectivity is the edge. Not prediction accuracy and selectivity.
The survivors shared a common trait: they were highly selective, only engaging when regime conditions matched their design parameters. Selectivity is the edge. Not prediction accuracy and selectivity.
Autonomous Intelligence For The Next Era of Finance
2026 © Otonomii LTD. All rights reserved.
TOP
Autonomous Intelligence For The Next Era of Finance
2026 © Otonomii LTD. All rights reserved.
TOP
Autonomous Intelligence For The Next Era of Finance
2026 © Otonomii LTD. All rights reserved.
TOP