OCM Brain
Service-backed AI layer for lead scoring, consent checks and sales-facing insight generation.
In production
Run commercial scoring and decision support inside stable backend workflows.
Engineering Notebook
Last updated
Sergio Ballesteros
Software Developer
Java · Spring Boot · AWS · Kubernetes
I build production systems.
Backend · Data · Artificial Intelligence
Operating backend, data and AI systems for commercial workflows in the energy sector.
Systems in Production
Systems I have to keep useful after deployment, not just finish once.
How I can help
I focus on practical systems that solve measurable business problems — not technology for technology's sake.
Web Analytics & Digital Infrastructure
Analytics, conversion tracking, Search Console, structured data, technical performance and web integrations.
Data & Automation
Turn repetitive processes and disconnected information into automated workflows and usable data.
AI Integration
Integrate AI into real business processes: document processing, internal tools, classification, assistants and decision support.
Custom Software
Build internal tools, APIs, dashboards and systems around specific business requirements.
Engineering Decisions
Why these systems were built this way, and what that choice bought in operation.
Why Spring Boot?
- Context
- Long-lived services needed predictable structure once they moved beyond the first delivery cycle.
- Decision
- Standardize core backend services on Spring Boot so service contracts, testing and operations stay consistent.
- Tradeoffs
- Heavier than minimal frameworks, but cheaper to own once the service becomes a real dependency.
Why PostgreSQL?
- Problem
- Reporting, workflow state and relational data were all part of the same operational surface.
- Decision
- Use PostgreSQL where transactions, ad hoc querying and clean relational modeling matter at the same time.
- Result
- Fewer compromises between business workflows and operational reporting.
Why Kubernetes?
- Context
- Several services had to be deployed, rolled back and observed in a repeatable way.
- Decision
- Run production services in Kubernetes so runtime behavior stops depending on one-off environment handling.
- Tradeoffs
- More platform complexity up front, less deployment ambiguity later.
Why OpenAI instead of local models?
- Problem
- Business workflows needed usable language capability before local-model ops were worth the cost.
- Decision
- Use hosted models where delivery speed, output quality and operational fit mattered more than ownership of the inference stack.
- Tradeoffs
- External dependency, but a faster path to production behavior that people could actually use.
Evidence Trail
Architecture, merge requests, deployment and monitoring as one engineering lifecycle.
System shape first
Define service boundaries, data ownership and failure paths before implementation starts.
- Inputs and outputs defined per service.
- Async work split from synchronous user flows.
- Storage chosen around operational queries.
Intent written down
Every change should explain what moved, why it moved and what operational risk it changes.
- Scope reduced to one production concern at a time.
- Tradeoffs captured near the implementation.
- Review driven by runtime impact, not just style.
Release as a controlled step
Containers, environment boundaries and rollout behavior are treated as part of the feature.
- Docker image defines the runtime contract.
- Kubernetes makes scaling and rollback explicit.
- AWS services stay behind predictable interfaces.
Production decides truth
Logs, import failures and AI validation loops decide whether the system is actually healthy.
- Real incidents feed the next engineering change.
- Data quality issues are treated as product issues.
- Operational fixes become documented system behavior.
- Architecture notes: service boundaries, queues and storage.
- Implementation notes: Spring Boot, Python and API contracts.
- Data evidence: SQL cleanup, validation and batch processing.
- Operational evidence: deployments, fixes and workflow hardening.
- The systems were built to run, not just to demo.
- Technical choices were made under operational constraints.
- Automation replaced repeated manual work.
- AI was integrated inside real business workflows.
Engineering Journal
Not a blog. A logbook of decisions, migrations, fixes and lessons learned.
Why I migrated this repository
Documenting the point where code organization stopped scaling with the system.
Fixing encoding issues while importing 900,000 records
How production-grade data cleanup differed from a quick script.
What I learned integrating Anthropic
Operational lessons from moving LLM workflows into real business processes.
Why I chose Spring Boot
Decision-making under production constraints instead of framework fashion.
Why I turned pricing logic into a platform
How repeated commercial analysis became a reusable system instead of analyst-owned work.
Why OCR accuracy was not enough
Why trusted structured payloads mattered more than promising extraction percentages.
Public Proof
Public evidence you can inspect or use: an interactive system demo, a live production product and an open backend architecture.
Energy Simulator
Turn an electricity bill into structured data, review the inputs and compare calculated scenarios. Try the complete workflow with a fictional invoice in about a minute.
Invoice → Extraction → Human review → Simulation → Results
Simplified public demo. Extraction is mocked; sample data, tariffs and prices are synthetic.
AI-assisted B2B prospecting engine: structured LLM research pipeline, deterministic opportunity scoring, layered domain/application/infrastructure backend.
layered backend architecture with evolving schema migrations and real tests
Full-stack product running in production. Real tests, CI/CD and a deploy pipeline that backs up the database before every migration.
safe, tested and reversible production deploys
Have a business problem that could be solved with software, data or AI?
I design and build practical digital systems — from analytics and automation to custom software and AI integration.