Platform / Barion Core
THE SYSTEMBENEATHTHE PRODUCTS.
Barion Core is the shared architecture used to build and operate every Barion AI product. Four layers carry a decision from raw sources to a recorded outcome, and a fifth concern runs across all of them.
- Layers
- Four, plus cross-cutting
- Model strategy
- Model-agnostic
- Running in production
- PSX Invest
Architecture
Four layers, driven by you rather than by scroll
Select a layer to see what it takes in, the mechanisms inside it, and what it hands to the next one. Every layer's detail is on this page whether or not you interact with it.
Assemble everything the decision depends on, from wherever it currently lives, and turn it into context a reasoning layer can use.
Takes in
- Structured records and tables
- Unstructured documents and notes
- External APIs
- Event and message streams
- Product-specific sources
Produces
- A single assembled context for one decision
Mechanisms
- Connectors
- Per-source adapters that pull at the rate the source produces, rather than on a fixed schedule that drifts out of date.
- Retrieval
- Selecting the subset of available information that is relevant to this specific decision, instead of passing everything forward.
- Context construction
- Normalising and assembling the retrieved material into a consistent shape, so downstream reasoning is not re-interpreting each source.
Generate the realistic options, evaluate them, and rank them. This layer produces candidates and comparisons, not a single opaque answer.
Takes in
- Assembled context
- Domain rules
- Objectives and weightings
Produces
- Ranked options
- Confidence and uncertainty
- Reasoning trace
Mechanisms
- Model orchestration
- Multiple models coordinated for different sub-tasks. The architecture is model-agnostic by design, so no single external provider is load-bearing.
- Rule engines
- Deterministic domain logic for the parts that must not be probabilistic, running alongside the models rather than after them.
- Ranking and optimisation
- Ordering options against the stated objective, so the output is a comparison a human can interrogate.
- Scenario generation
- Showing how an option behaves under different conditions, which is what makes risk discussable.
Decide whether the proposed output is allowed to leave. Output that fails a policy is held for review rather than shipped.
Takes in
- Ranked options
- Reasoning trace
- Policy set for this product
Produces
- Cleared output
- Held-for-review items
- Escalations
Mechanisms
- Permissions
- What this system may read, propose and do, expressed as boundaries rather than as instructions in a prompt.
- Constraints
- Hard limits that cannot be argued past. Risk ceilings, restricted actions, domain-specific prohibitions.
- Validation
- Checking the proposed output against the constraints and the context it came from before anyone sees it.
- Human approval
- A required signature on the critical path, with the approver and the timestamp recorded, not an optional review step.
- Escalation
- A defined route to a named human when confidence is low, a constraint is hit, or the situation was not anticipated.
Get the cleared result to the place the decision is actually made, in the form that place expects.
Takes in
- Cleared output
- Approval record
Produces
- A decision made, with the basis for it recorded
Mechanisms
- Dashboards and product interfaces
- Where a person reads the options and decides. PSX Invest is this layer for the markets domain.
- Alerts
- Pushing the decision to the human when the trigger fires, rather than waiting for them to come looking.
- APIs and webhooks
- For systems that consume the intelligence programmatically and build their own surface on top.
- Workflows and agents
- Where a bounded action is executed on the other side of an approval gate.
Across every layer
Cross-cutting
These are not a fifth stage at the end. They run alongside all four layers, which is what makes a decision reviewable months later.
- Logging
- What arrived, what was concluded, which rules applied, who approved.
- Observability
- The live behaviour of the system while it is running, not after an incident.
- Evaluation
- Whether output quality is holding, measured against defined criteria.
- Monitoring
- Health, latency and error behaviour of the system in production.
- Feedback
- What actually happened after the decision, fed back into the next one.
Trade-offs
What production intelligence actually costs
There is no configuration that is fast, cheap, maximally accurate, fully explainable, perfectly safe and free of human involvement. Every real system picks a position, and the honest thing is to say which.
- 01Speed against deliberation
- More reasoning passes usually mean better options and a slower answer. A trading scan can spend minutes after close. An operational alert cannot. The budget is set per product, not globally.
- 02Cost against depth
- Every additional model call has a price. Depth is spent where the decision is expensive and withheld where a cheap deterministic check is sufficient.
- 03Accuracy against explainability
- The most accurate arrangement is not always the most explainable one. In domains where a decision must be defended, we accept a less clever system that can show its work.
- 04Safety against usefulness
- Tighten constraints far enough and the system refuses everything, which is safe and worthless. Constraints are set at the point where the remaining output is still useful.
- 05Human involvement against throughput
- Approval gates cost time. They are placed where the consequence justifies the delay, rather than everywhere or nowhere.
- 06Autonomy against reversibility
- How much a system may do alone is governed by how easily its actions can be undone. Irreversible actions do not get autonomy.
Integration
How Barion Core reaches your systems
Patterns rather than a connector catalogue. What exists for a given deployment is scoped during design, because claiming a library of pre-built integrations we have not written would be dishonest.
- APIs
- Request and response access to the intelligence layer for systems that build their own surface on top of it.
- Webhooks
- Push delivery when a condition is met, so the receiving system does not have to poll for changes.
- Data pipelines
- Scheduled or event-driven ingestion from the sources a decision depends on, at the cadence the source supports.
- Product interfaces
- A working surface where a human reads options and decides. PSX Invest is this pattern for markets.
- Approval systems
- Routing a required signature to the person who holds the authority, and recording what they approved.
- Reporting layers
- Evidence and outcome data pushed into whatever the organisation already uses for review and audit.
Deployment
How a system gets from design to operating
The architecture is only half of it. The other half is the sequence that gets a system into production without discovering the constraints on the way.
- 01Decision first, data second
- The decision being improved is defined before anything is built, including who owns it and what a better outcome looks like.
- 02Constraints written down
- What the system may never do is agreed and encoded early, because retrofitting boundaries into a working system is how they end up incomplete.
- 03Pilot against real data
- Narrow scope, real inputs, guardrails already in place. A pilot on synthetic data proves very little.
- 04Run alongside before replacing
- Comparing the system’s behaviour with the current process is what turns a demo into evidence.
- 05Operate and watch for drift
- Once running, behaviour is monitored against expectation. A system nobody is watching is a system nobody can trust.
Deployment requirements around residency, retention and access are assessed during deployment design with the organisation involved. What we can and cannot commit to today.
Continue
Where to go next
- Build with BarionThe structured ways an organisation can use this architecture for its own problem.
- PSX InvestThis architecture in production, with the boundary between core and product spelled out.
- Governance and observabilityThe cross-cutting layer in practical detail: controls, evidence and override.
Build on Barion Core
If you own a repeated decision that has to be explainable and bounded, this architecture is the starting point rather than a greenfield build.