Platform architecture

One platform. Five agents.

Our products are not five separate tools that happen to share a logo. They are five agents on one runtime, one data plane and one assurance layer — which is why the second deployment inside an institution costs a fraction of the first.


The stack

Read it top to bottom.

Work enters at the top, passes a human approval gate, is executed by an agent, and lands in your systems of record. Everything below the gate runs inside your own tenant.

PAAR Systems agentic platform architecture Seven layers, top to bottom: people and triggers; a human approval gate; the five agents Passage, Loom, Redress, Attune and Relay; the agent runtime with planner, action router, memory and cost controls; shared platform services covering retrieval, guardrails, evaluation, observability and a swappable model layer; the data plane ingesting documents, email, transcripts and core tables with entity resolution and lineage; and the institution’s systems of record. Everything below the approval gate sits inside a dashed boundary marked as the customer tenant with no data egress. YOUR TENANT · VPC OR ON-PREM · NO DATA EGRESS PEOPLE & TRIGGERS Analysts & ops teams Contact-centre agents Programme & risk owners Scheduled / API triggers HUMAN APPROVAL GATE Nothing consequential leaves the platform without a named human approving it Four-eyes, delegations and segregation of duties enforced · agents draft, assemble, rank and flag — they do not decide ESCALATE AGENTS PAAR Passage Migration Reads legacy code Extracts business rules Parallel-run equivalence PAAR Loom Data processing Docs, email, transcripts Entity resolution Field-level lineage PAAR Redress Remediation Cohort reconstruction Auditable recalculation Evidence packs & letters PAAR Attune Voice Real-time agent assist Call summarisation Every call reviewed PAAR Relay Workflow Process orchestration Exception routing Approvals enforced AGENT RUNTIME Planner & task graph Decomposes work into steps Tool & action router Only allow-listed actions Working memory & state Case context, checkpoints Cost, latency & limits Budgets and rate control SHARED PLATFORM SERVICES Retrieval Vector + document store Policy & procedure corpus Guardrails & policy PII redaction Prompt and output filters Evaluation Golden sets, regression gates Live quality monitoring Observability & audit Immutable action log Model & prompt versions Model layer Frontier + local models Swappable, no lock-in DATA PLANE Documents & forms PDF, scans, statements Email & transcripts Threads, call audio Core system tables Batch and CDC feeds Resolution & lineage One customer, one truth SYSTEMS OF RECORD Core banking Lending & originations CRM & servicing Data warehouse Case, BPM & telephony

Scroll horizontally on a narrow screen. Every box is a real component, not a marketing layer — ask us to walk through any one of them.

Layer by layer

What each band actually does.

People & triggers

Work arrives from an analyst, a contact-centre agent, a programme owner, or a schedule. Agents are not a chat window bolted onto the side — they sit in the workflow where the work already starts.

Human approval gate

The single most important box on the page. Agents draft, assemble, rank and flag; a named human approves anything that touches a customer, a ledger or a regulator. Four-eyes, delegations and segregation of duties are enforced by the platform, not by convention. Agents can escalate upward at any point when confidence is low or a case falls outside their remit.

The five agents

Passage, Loom, Redress, Attune and Relay. Each owns one domain and can be bought on its own, but they share everything below this line — which is why the second one costs a fraction of the first to deploy.

Agent runtime

The planner decomposes a request into steps and keeps a task graph. The action router will only call tools on an allow-list agreed with you. Working memory holds case context and checkpoints so a long-running job can be paused, inspected and resumed. Cost, latency and concurrency budgets are hard limits, not dashboards.

Shared platform services

Retrieval over your documents and policy corpus. Guardrails for PII redaction and prompt and output filtering. Evaluation with golden datasets, regression gates before release and live quality monitoring. Observability that writes an immutable log of every input, model version, prompt version, output and reviewer action. And a swappable model layer, so frontier progress is an upgrade rather than a rebuild.

Data plane

Loom doubles as the platform's ingestion layer: documents, forms, email, call audio and core system tables, joined and resolved to one customer with field-level lineage. Any number a downstream agent produces can be traced back to the page and record it came from.

Systems of record

Core banking, lending and originations, CRM and servicing, the warehouse, and case, BPM and telephony platforms. The platform integrates with what you run; it does not ask you to replace it.

Why a platform

The shared substrate is the product.

01

Approve once, reuse

The second agent inherits the security review, the data flows, the model risk documentation and the deployment topology that the first one cleared. Most of the cost of AI in a bank is the approval, not the build.

02

One audit trail

Every agent writes to the same immutable log in the same shape. An examiner asks one question and gets one answer, rather than five vendors’ exports stapled together.

03

One data plane

Documents extracted for a remediation program are already available to the credit or workflow agent, resolved to the same customer, with the same lineage.

04

Start with one

Nothing here requires a platform purchase. Buy the agent that solves your worst problem; the substrate comes with it and the next one is faster.