Home / Blog / How Duco transforms financial data automation with Amazon…
Tech News

How Duco transforms financial data automation with Amazon Bedrock AgentCore

Duco rebuilt its banking operations platform on Amazon Bedrock AgentCore, slashing reconciliation setup times 10x while isolating 20 billion monthly records.

By Dillip Chowdary • Nov 17, 2026 • Source: AWS Machine Learning Blog

How Duco transforms financial data automation with Amazon Bedrock AgentCore
Image: AWS Machine Learning Blog

At a glance

Product
Amazon Bedrock AgentCore
Monthly volume
20 billion transactions
Client base
200+ financial institutions
Onboarding savings
70 percent cost reduction
Speed improvement
10x faster deployment
Context reduction
90 percent less overhead

Duco rebuilt its financial data automation platform on Amazon Bedrock and Amazon Bedrock AgentCore, replacing classical machine learning systems with an isolated multi-model agent architecture. The enterprise deployment automates mission-critical reconciliation and extraction workflows across banking institutions without cross-tenant data exposure, according to AWS Machine Learning Blog's report. The modernization reduced end-to-end reconciliation deployment time by ten times, lowered client onboarding expenses by 70 percent, and decreased operational context overhead by 90 percent.

This architectural review details how financial infrastructure engineering teams deploy production AI agents within strictly regulated banking perimeters. Systems architects, compliance officers, and data operations leaders will see how tenant-scoped runtime gateways, containerized model context protocol services, and tiered model orchestration satisfy strict enterprise isolation demands. The implementation demonstrates how tier-one capital markets institutions transition from brittle historical predictive models to resilient autonomous reasoning systems.

01Classical machine learning operational limits

Duco operates as an enterprise data automation platform serving tier-one global financial institutions, processing 20 billion transactions every month across more than 200 clients. The customer base includes 7 of the top 20 banks and 10 of the top 20 asset managers, supporting over 20,000 automated operational workflows. Traditional systems depended on predictive machine learning classifiers for pattern recognition and supervised record labeling across ingestion pipelines.

Those legacy statistical models required massive volumes of historical training data, performed poorly during unexpected market disruptions, and required costly manual customization for new client formats. Fixed mathematical algorithms could not adapt to rapidly evolving trade confirmation structures, ambiguous counterparty data feeds, or changing settlement rules across distributed ledgers. Engineering teams spent excessive cycles retraining bespoke supervised pipelines whenever enterprise file layouts or regulatory reporting standards shifted.

Rob Gates, chief architect and innovation officer at Duco, pointed out the structural ceiling that classical algorithmic techniques imposed on institutional scale. Gates noted that the classical prediction models worked to an extent, but they were not agile or flexible enough for unpredictable banking workloads. He explained that traditional machine learning models can only amplify what they have already learned, leaving them unable to assist with subsequent operational steps or detect novel structural anomalies.

02Managed foundation runtime selection criteria

Duco selected Amazon Bedrock to secure managed API access to high-performance foundation models alongside Amazon Bedrock AgentCore to execute agent workloads. The infrastructure architecture ensures sensitive customer financial data remains strictly within protected cloud environments without ever being repurposed for model training routines. Private VPC endpoints connect internal systems directly to Amazon Bedrock, preventing raw transaction traffic from traversing public networks.

The core runtime migration preserved existing software assets because Duco had already deployed production agent pipelines using LangChain and LangGraph frameworks. Amazon Bedrock AgentCore provides native runtime compatibility with these open-source primitives, allowing existing execution graphs and agent loops to function without code refactoring. The managed control plane eliminated custom agent host fleet maintenance, shifting infrastructure patching, agent lifecycle operations, and scaling requirements directly to AWS.

Unified model routing through a single managed API gave engineers instant access to multiple model families, including the Amazon Nova 2 family, Anthropic, OpenAI, Cohere, Meta, and Mistral AI. Platform administrators switch active models for specific agent tasks by modifying runtime configuration parameters rather than rewriting orchestration logic. This standardized foundation environment met the governance, auditing, and compliance frameworks required by regulated financial entities.

03Dedicated cloud infrastructure isolation layers

Advertisement

Tech Pulse Daily

Get tomorrow's pulse first

Join engineers who read Tech Pulse before stand-up. Free, weekday mornings.

Duco enforces a strict multi-tenant SaaS architecture where every institutional client receives dedicated, unshared cloud resources across all storage and database tiers. Workloads run inside Amazon Elastic Kubernetes Service clusters segregated through tenant-specific namespaces and strict Kubernetes network policies. Compute autoscaling operates per tenant, ensuring that sudden high-volume processing spikes from one bank never degrade adjacent client performance.

Data tier isolation is maintained by provisioning dedicated AWS resources for every customer account rather than relying on logical row-level tenant tagging. Each client deployment receives an isolated Amazon Aurora MySQL database, dedicated Amazon S3 buckets, distinct AWS Key Management Service encryption keys, and an independent Amazon ElastiCache cluster. Tenant messaging and identity stacks are partitioned using dedicated Amazon SES and Amazon SQS queues alongside isolated Amazon Cognito user pools.

The architectural boundary prevents any cross-tenant data exposure at rest, in transit, or during compute cycles. Network routing keeps inter-service communication contained within dedicated private subnets governed by VPC security groups. By eliminating shared persistent storage across clients, the infrastructure complies with global banking regulations that prohibit commingling regulated customer records.

04Gateway authentication and protocol mediation

Security enforcement relies on continuous identity propagation across every infrastructure layer before agent reasoning logic executes. Each inbound transaction carries a JSON Web Token issued exclusively by the customer's dedicated Amazon Cognito user pool. The inbound authorizer of the Amazon Bedrock AgentCore runtime validates this identity token, ensuring untrusted external requests cannot reach agent memory or trigger tool calls.

AI agents access operational tools exclusively through a tenant-dedicated AgentCore Gateway running inside private Amazon VPC network mode. The gateway routes requests to containerized Model Context Protocol services hosted in the tenant's isolated Amazon EKS namespace, cutting context overhead by 90 percent. MCP servers validate both the invoking agent and the authenticated user identity against Cognito, guaranteeing that every downstream tool execution inherits proper user authorization.

Every tenant runs an independent AgentCore runtime and gateway instance bound to its unique Cognito pool, preventing shared context or cross-tenant reasoning leakage. Built-in session isolation ensures that proprietary schemas, internal data dictionaries, and institutional accounting rules remain confined to that single institution. Private AWS network endpoints keep all agent traffic isolated from public transit routes.

05Multi-model delegation and agent tiers

Duco structures agent responsibilities across distinct operational tiers, pairing specific model capabilities to individual workload requirements to optimize accuracy and operating costs. A planning agent analyzes incoming unstructured exceptions and formulates multi-step execution plans before downstream execution agents trigger backend actions. Specialized sub-agents then resolve transaction mismatches, interpret complex trade confirmations, or direct lightweight query traffic.

| Model | Primary role | Use cases | | Anthropic Claude Opus | Planning agent | For complex reasoning and understanding, formulating up front plans for other agents to execute on | | Anthropic Claude Sonnet | Main reasoning & complex decision-making | Exception investigation, reconciliation logic, rule interpretation | | Anthropic Claude Haiku | Lightweight tool routing decisions | Fast routing |

The tiered multi-model strategy delivers significant operational efficiencies compared to legacy manual or static machine learning onboarding pipelines. Reconciliation deployment time dropped by a factor of ten, allowing institutions to launch complex data pipelines in hours instead of weeks. Client onboarding expenses fell by 70 percent, proving that managed agentic architecture can scale safely across heavily regulated capital markets.

Frequently asked questions

What models does Duco run inside Amazon Bedrock?

Duco uses Anthropic Claude Opus for planning, Claude Sonnet for complex reasoning, Claude Haiku for fast tool routing, and can access the Amazon Nova 2 family, Meta, Mistral AI, Cohere, and OpenAI.

How does the platform maintain tenant isolation for banks?

The architecture provisions dedicated Amazon Aurora databases, Amazon S3 buckets, AWS KMS keys, Cognito user pools, and isolated Amazon EKS namespaces with dedicated AgentCore runtimes for each tenant.

Does Duco share agent memory across financial institutions?

No, each client runs an isolated AgentCore runtime and gateway with session boundaries that prevent shared context, memory, or tooling across institutions.

Sources

Dillip Chowdary

Author

Dillip Chowdary

Writes Tech Bytes coverage of AI, engineering, and the tools that actually ship. Editor of Tech Pulse Daily.

Related on Tech Bytes

Advertisement

5-min tech signal

Weekday briefing for engineers who skip the noise.

No spam · Unsubscribe anytime

Advertisement

✈️ CareerPilot

Your AI job-search copilot

Match your resume against live Ashby, Greenhouse & Lever openings — fit scores, job-specific resume optimization and email alerts.

Find matching jobs →

Free Tools

Browse all tools →