Empowering Real-Time Payments Intelligence
A migration to Amazon Bedrock AgentCore
Deuna is a Latin American fintech platform for payments orchestration, one-click checkout and fraud prevention. Processing millions of transactions for enterprise clients like KFC and Sony, it connects over 400 global payment providers into a single API.
Their core innovation is Athia, a “commerce-aware” AI platform that turns fragmented payment data into real-time growth strategies through a natural language interface.
Explosive adoption, and a multi-agent architecture that could not keep up
Athia grew 24% monthly, with over 50 enterprise clients using or piloting it — and the scaling exposed a bottleneck: Deuna’s multi-agent architecture (8+ agents) took 1.5 to 2 minutes to answer a complex query.
Deuna engaged binbash for GenAI architecture, DevOps engineering and technical leadership: untangle the bottlenecks, cut latency by 50%+, and build a secure, cost-effective AWS-native foundation.
A Guided Evolution Model, not a hand-off
As a strategic advisory partner we implemented a “Guided Evolution Model” built on knowledge transfer and architecture design: Deuna’s engineering team built the solution, and we guided them through migrating their AI orchestration workload to Amazon Bedrock AgentCore.
Israel GerardoDirector of Data, DEUNA 5.0 verified review on Clutch“They had an in-depth knowledge of AWS tools and experience in real-world projects.”
Acting as CTO, Product Management and Cloud Architecture
Acting with CTO, Product Management and Cloud Architecture capabilities, we mentored Deuna’s team through the modernization:
GenAI Architecture & Mentorship: We advised Deuna’s engineers on moving their multi-agent workflows to Amazon Bedrock AgentCore, recommending Anthropic models like Claude Sonnet and Claude Opus to speed up reasoning and cut latency.
Security & Compliance Advisory (PCI DSS): We designed an AI safety layer on Amazon Bedrock Guardrails, with reference implementations, and guided Deuna on configuring it to scrub PII and block prompt injection before they reach the LLM.
Observability & Evaluation Strategy: We co-designed an automated “LLM-as-a-Judge” pipeline on CloudWatch, Kinesis Firehose, Lambda and Amazon Bedrock Evaluations, mentoring the team to run continuous accuracy monitoring themselves.
A team that can extend it without us
Deuna’s engineering team built a secure, high-performance AWS-native foundation on our architectural guidance, decoupling from expensive third-party compute, taking control of their GenAI operating costs and targeting sub-minute response times.
- 50%+
- Target performance lift in chatbot response times
- <60s
- Multi-agent query SLA (down from 1.5–2 min)
- 10–15x
- ROI delivered for enterprise merchants
Delivered in four milestones
Internal Assessment & Discovery: Audit the data/AI bottlenecks and identify where latency can come out.
Architectural Design: Co-design the AWS-native target architecture and the blueprint for an Amazon Bedrock AgentCore Proof of Concept.
Guided Implementation & Acceleration: Deuna’s team builds the PoC, with binbash running weekly sprint planning, code reviews and workshops.
Security & Evaluation Hardening: Advise on PCI DSS-compliant PII scrubbing via Bedrock Guardrails and an automated semantic evaluation pipeline.
From a scaling crisis to a self-sufficient GenAI team
binbash guided Deuna in evolving “Athia” from a constrained, high-latency application into a scalable GenAI product, working strictly to the AWS Well-Architected Framework and delivering technical advisory, code reviews and architectural blueprints. Deuna’s own developers implemented Amazon Bedrock AgentCore and its guardrails, which solved the immediate scaling crisis and left them a self-sufficient engineering team.