European RegTech Platform

The Challenge

Manual processing was slow, error-prone, and didn’t scale. Handoffs between teams caused delays and information loss. Regulatory audits were painful because the full processing history was scattered across email threads, spreadsheets, and disconnected systems.

The company needed intelligent automation that could handle complex multi-step workflows end-to-end, make auditable decisions at each stage, recover from interruptions (cases can take weeks to resolve), and integrate with multiple backend systems through a controlled interface.

Our Solution

We designed and built a multi-agent orchestration platform using LangGraph’s supervisor pattern. A central orchestrator agent routes incoming cases to five specialized sub-agents, each responsible for a distinct domain:

• Assessment Agent — evaluates case details against historical data and generates cost/risk assessments
• Coordination Agent — manages external service provider assignment, scheduling, and status tracking
• Payment Agent — handles payment authorization, disbursement tracking, and reconciliation
• Recovery Agent — manages long-running recovery and recoupment processes (spanning months)
• Compliance Agent — validates case data against regulatory requirements, coverage rules, and policy limits

Agents communicate through Pydantic v2 strict data contracts validated at every boundary — no silent data corruption, no malformed handoffs. The supervisor doesn’t do domain work; it coordinates, tracks state, and handles escalation.

State Management

All agent state is checkpointed to PostgreSQL after every graph node execution. Cases resume from last checkpoint on failure. The full state history is queryable — every decision, every transition, every agent output is recorded. This is what makes regulatory audits tractable: the audit trail isn’t reconstructed, it’s the primary data structure.

Integration Layer

An MCP (Model Context Protocol) server gateway decouples agent internals from external systems — document management, regulatory databases, payment systems, provider networks. Tool-level access control lives in the gateway, not scattered across agent code. External systems with legacy protocols are abstracted behind clean tool interfaces.

Multi-Modal Intake

Cases enter via three channels — LiveKit-based voice processing (real-time transcription with interactive clarification), email parsing (structured field extraction from attachments and body), and web portal chat. All channels converge on the same supervisor for consistent processing.

Observability

MLflow tracing across full agent execution chains tracks latency, token usage, correctness, and cost per case. 15+ end-to-end test scenarios validate the complete case lifecycle — not synthetic tests, but representative real-world scenarios.

Infrastructure

AWS (S3, VPC, KMS per tenant), Databricks with Unity Catalog for data processing, Terraform/Terragrunt for multi-region deployment across a Region, Domain, and Environment hierarchy. 22-component suite deployed via GitOps. Datadog for infrastructure monitoring.

What Didn’t Go Smoothly

The platform is transitioning from monolithic to service-extracted architecture — phase 2 of a 3-phase decoupling roadmap. Individual high-latency agents are being extracted as independent services with API boundaries. Full event-driven microservice decomposition is designed but not yet in production. This is an honest reflection of where production agentic systems are today: the architecture is sound, but the operational maturity is still evolving.

Results

• Automated routing and processing of regulatory compliance cases through supervised agent orchestration
• Complete audit trails satisfying financial regulatory requirements — every decision traceable
• Manual review reduced to exception cases and human-in-the-loop approval gates (e.g., payments above threshold)
• Multi-modal intake consolidated three separate manual processes into one consistent pipeline
• Platform designed for multi-region, multi-tenant deployment with tenant-level data isolation (bucket-per-tenant, KMS encryption per tenant)

Technologies Used

LangGraph, Python, Pydantic v2, PostgreSQL, MCP, MLflow, LiveKit, Databricks, Unity Catalog, Terraform, Terragrunt, AWS, Datadog, GitOps