Four interlocking layers — Data, Model, AI and LLM — bound by a governance and lineage spine that runs end to end. Built ground-up for the way banks and insurers actually ship work the regulator can defend.
The Wersel Stack is the reference architecture our practitioners use to ship Data, Model, AI and LLM workloads inside a regulated bank or insurer. It is opinionated about layering, lineage and governance — and deliberately neutral about specific products, because the right vendor in your environment is rarely the same as the right vendor in someone else’s.
What it does fix is the shape of the work: which capabilities sit at which layer, how the governance spine cuts through every tier, and where the seams are that audit, model risk and the supervisor will inspect. Use it as a planning canvas for your next regulatory milestone, or as a target architecture for an estate that has grown faster than its operating model.
The lifecycle view follows a single decision from ingest to enforce. The layer view shows the engineering deliverables that ship at each tier — Data, Model, AI and LLM — under one governance spine.
A reasoned, traceable and governable fabric for Data, Model, AI and LLM workloads — wired into the banking and insurance domains on one side, and the people who answer to the regulator on the other.
The central reasoning fabric: aggregates upstream signals, generates the explanation graph and surfaces gap-detection telemetry to monitoring loops — so every decision can be reconstructed end to end.
Each layer ships with its own engineering deliverables and its own model-risk and lineage obligations. The spine ensures every input, every model and every output traces back to a controlled source — at the cadence your regulator now expects.
The ingestion, contract, warehouse and lineage primitives that turn source-system chaos into a versioned, governed, model-ready surface.
Typed contracts at the boundary, replayable ingestion, schema-evolution policy. The source-system surprises absorbed once, not every quarter.
A versioned analytical store with time-travel, branching and reproducible builds. The audit can re-run last quarter’s report against last quarter’s data.
Captured automatically at ingestion and transformation; CDE status carried as a live attribute, not a spreadsheet label.
Tokenisation, residency-aware routing and lawful-basis tags — engineered into the pipeline rather than bolted on at the report.
Where the classical risk, actuarial and pricing models live — engineered with the validation harness, challenger review and documentation the regulator will ask for.
PD / LGD / EAD, IFRS 9 stage transitions, macro-overlays and reverse-stress narratives — built with continuous lineage and explainable methodology bridges.
IFRS 17 measurement, Solvency II / UK reform, BPA pricing under PRA cashflow tests, Matching Adjustment ALM monitors.
Application scorecards, affordability matrices and personal-lines pricing with Consumer Duty fair-value evidence attached at the decision.
Pre-release validation, ongoing monitoring, documented challenger review and methodology bridge — generated alongside the model, not added before the meeting.
Production ML for fraud, claims, AML, marketing and operations — deployed under the same model-risk discipline as the classical credit and actuarial estate.
Versioned features, reproducible training runs, point-in-time correctness — so the model that shipped is the model the validation report describes.
Real-time scoring with calibrated thresholds, false-positive demographic monitoring and investigator-loop feedback wired into retraining.
Continuous evaluation against documented thresholds — bias proxies, drift, refusal behaviour and Consumer Duty signals all under live control.
Every model — vendor or in-house — carried as an inventory record with owner, purpose, tier, validation cadence and retirement plan.
Retrieval pipelines, agentic workflows and copilot surfaces — engineered with prompt/response lineage, evaluation harnesses and human-in-the-loop patterns the supervisor can inspect.
Document, policy and customer-data retrieval with provenance preserved — grounding the model on a controlled corpus rather than its training set.
Bounded agents that decompose work into auditable steps, with tool allowlists, blast-radius limits and step-level evaluations.
Context window, tool calls, model outputs and downstream actions captured per interaction — the audit trail generative AI was missing.
Hallucination, accuracy, refusal and Consumer-Duty evaluations on live traffic, with tier-based escalation to a human reviewer when confidence drops.
A continuous control fabric that cuts vertically through Data, Model, AI and LLM — so every change has a controlled cause and every output a defensible source.
The same four layers, the same spine — instantiated against the regulatory reality of each vertical. These are the targets our embedded teams ship against.
Talk to our practitioners about adopting the Wersel Stack as a target architecture — or commissioning a current-state assessment of your Data, Model, AI and LLM estate.
Or write to us directly: hello@wersel.io
