Juan A.

Most AI features don't fail because of the model. They fail because the backend was never designed for non-deterministic systems. I've spent 12 years building Java and Spring platforms for banking, telecom, healthcare and retail — systems where an outage is an incident report, not a bad day. For the last two years I've been putting LLM agents inside that same class of system and keeping them running. That is the work I do: AI inside platforms that already have auth, audit trails, PII rules, SLAs and someone on call. What that looks like: — Digital banking on Backbase DBS: arrangements, transactions, limits and maker-checker approval matrices. Spring Cloud Gateway, OAuth2/OIDC, saga patterns for money flows that cannot end half-done. — A five-service telecom platform on GCP implementing TMF standards. Cloud Run, Spanner with SCD-2 history tables and a transactional outbox, Pub/Sub with dead-letter queues, customer PII tokenized through Cloud DLP with versioned HMAC keys, secrets moved out of environment variables into Secret Manager. — An agent platform I built and still operate: WhatsApp and voice agents on AWS Lambda, DynamoDB, EventBridge and SQS FIFO. Real constraints, not a sandbox — per-portfolio messaging limits, template category audits, error codes you must never retry, a one-second turn budget on voice, and a cost per conversation that has to stay below what the customer pays. How I work: I treat an agent like any other distributed system. Every prompt or model change ships as an experiment: a regression bank of replayed real conversations, canary promotion, and a config fingerprint on every logged run — because "the prompt got better" is not a claim you can defend without one. I use AI to write code and I review every line of it at the boundaries that matter: schemas, money, auth, and anything that deletes. Those it does not touch unreviewed. I work in English with distributed teams across the US and Latin America, from Bogotá. If you are adding AI to a platform that already has customers, compliance and an on-call rotation, that is the problem I solve. 12 years of enterprise backend development. I've built: → A loan processing engine for a banking platform — retail and corporate lending workflows, credit decisioning APIs, multi-stage approval pipelines → A custom eSign web application — full design and development, then seamless feature expansion for a returning client → A serverless inventory management system on AWS (Lambda + DynamoDB + API Gateway + Cognito) → Clinical documentation systems handling care gap detection across 31 medical categories I specialize in the kind of backend work where getting it wrong is expensive: financial services, healthcare, compliance-heavy domains. Systems where you need someone who thinks about edge cases before they become production incidents. Tech stack: Java 8/11/17 · Spring Boot · Spring Security · Spring Batch · Spring Data JPA · Hibernate · PostgreSQL · MySQL · Oracle · DynamoDB · Apache Kafka · Docker · REST APIs · AWS (Lambda, EC2, ECS, S3, Step Functions, API Gateway) · OAuth 2.0 · JWT · CI/CD (Jenkins, SonarQube) · 90%+ test coverage (JUnit 5, Mockito) 100% Job Success. Every project delivered on scope. Every client rated me 5 stars on communication and quality. Based in Colombia (EST-aligned). Fluent English and Spanish. Available 30+ hrs/week. I respond within hours, not days. Looking for a Java backend developer, Spring Boot specialist, or AWS microservices engineer? Let's talk about your project.