ARCHITECTURE

Built to be operated by people and software alike.

Every capability in PhynRisk and PhynRate is an operation that a person, a script or an AI agent can call — on one deterministic engine that records how every number was made.

See it running The four design decisions ↓

THE STACK

Agents need something underneath them.

Agents — people or software

The same operations, whoever is driving.

The operating surface

Every capability a typed operation with a published schema.

The measures

Capital, ECL, IRRBB, liquidity and ratings, all calculated from one engine.

The engine

One deterministic calculation over one set of cash flows.

Your data

Read through the same operations, not replaced.

AI is the top layer, not the foundation. The numbers a supervisor relies on come from the layers below it.

DESIGNED IN, NOT BOLTED ON

Four decisions made before the first line of code.

Everything is an operation

Every capability is a typed operation with a published schema, served over MCP (Model Context Protocol). Our own screens use the same operations — nothing hidden behind them.

Why it matters: an agent can do anything a user can.

The calculation flow is open

The order of calculation is configuration, not compiled code. An agent can see which rule runs at each step, and can compose a new flow.

Why it matters: you change the process without waiting for a software release.

Try, fail, correct

A single instrument or scenario runs in seconds, and failures come back as structured errors: what broke, which input, which rule.

Why it matters: an agent gets it right in minutes, not after an overnight batch.

Lineage recorded, not reconstructed

The engine records inputs, assumptions and intermediate values as it calculates.

Why it matters: lineage pieced together afterwards is an argument. Lineage recorded as you go is evidence.

None of these is a feature. Each is a decision that is easy at the start and out of reach afterwards.

WHAT IT MAKES POSSIBLE

Because the platform can be driven, agents can do the work.

Implementation

Map the book, set behavioural assumptions and configure the measures from plain-language intent, then put it into production under review.

Daily use

Read documents, fill in assessments, and calculate and explain results for the person doing the work.

Change

Apply a regulatory or business change without writing a specification and waiting in the queue.

Bring your own agents. If your bank has AI agents of its own, you can point them at the same operations. Ours are simply the first users.

FOR VALIDATORS AND AUDITORS

Deterministic where it counts.

What ran, in what order, on which inputs.

The open flow and the recorded lineage together give a complete explanation of any figure.

The model never calculates the number.

AI configures and explains. The calculation itself stays deterministic, reproducible and testable.

Sign-off is its own step.

The agent drafts; publishing is a separate operation a person performs, with the approval recorded.

Don't take our word for it. Check it.

In a technical session, we'll show you on the running system:

the operations and their schemas — and our own screens using them
the calculation sequence, read as configuration
one instrument run in seconds, and a structured error when an input is missing
any figure traced back to its positions, curves and parameters
an agent drafting a change, and a person publishing it

See it running

45 minutes with the people who built it. No slides.