01
AI automation
Workflow orchestration and integration glue between services, schedules and the agent layer — with an approval step in front of anything irreversible.
Automation approachEnrique Rubio
AI agent systems built to be safe with real access.
I build small, focused software systems — AI agents and automation with real approval gates and audit trails, and the data infrastructure underneath them — mostly for my own projects and for direct client work.
- years on production data platforms
- 7
- systems shown below
- 4
- bypass paths around approval gates
- 0
Capabilities
Four disciplines, one operating model
Each one has its own shape — a pipeline, a reasoning loop, a stream, a stack — and they all run through the same approval-gated core.
02
AI agents
Agents built on Claude Code that read state, decide and act on real infrastructure — inside identity and scope checks, with immutable artifacts and audit trails.
Agent architecture03
Data & systems
Seven years on production data platforms in banking and payments: Cloudera CDH/CDP, Kafka, hybrid on-prem/AWS, platform security and L3 support.
Background04
Product & software
Small, focused products shipped end to end and designed as layers I can reason about, operate and keep correct.
Selected systemsSelected systems
Real, running infrastructure
Three systems in active use, plus one still light on public detail. Full case studies — problem, constraints, architecture, decisions — live on Projects.
Uses YouTube API ServicesFeatured · content operations
Traffic OS
private · in production
A private content-operations application with a human approval gate before anything publishes — automation of the workflow, never the judgment call.
- 01CreativeProvenGeneration and review architecture is built.
- 02ApprovalHuman gateCreative approval, then a separate final approval. The approved artifact is hash-pinned.
- 03PublishProvenChannel identity checked, one-shot canary. One real public canary published.
- 04MetricsMeasuringCollected automatically for published videos.
- 05ExperimentMeasuringExperiment state is recorded against those metrics.
- 06DecisionWaiting for dataNeeds enough measurement windows. Nothing is decided automatically.
Auto-publication: disabledAny further publish needs a fresh human approval. Every publication is written to an append-only audit history.
AI agent operations system
personal · in production, human-gated
Personal orchestration infrastructure: autonomous agents act on real infrastructure, with a dedicated review UI for anything flagged as drift. It now includes a content pipeline: a real 1.8 GB long-form video job taken to a finished video, and a daily blog pipeline. A person approves everything before it publishes.
GlossAtlas
public · glossatlas.com
An independent cross-brand reference for paint sheen and finish equivalence — no brand affiliation.
CompatSense
in active use · limited public detail
Connectivity and compatibility intelligence tooling.
How it actually works
The operating model behind these systems
Every system above runs on the same shape: a human, a reasoning layer, a deterministic orchestration layer, an approval gate, and an audit trail. Hover or focus a node to see how it connects — the full text version is right below the diagram either way.
Map legend — every node and connection, in plain text
- Human
- Directs intent and makes the final call on anything irreversible — approves or denies each gated action.
- Claude / Claude Code — reasoning
- The reasoning layer: reads state, decides what to do, and drafts the action. Handles judgment, not deterministic glue.
- n8n — deterministic orchestration
- Runs scheduling, integration calls, and data movement between systems, and enforces the approval gate as a concrete workflow step an agent can’t quietly talk itself out of.
- Approval gate
- Anything irreversible — publishing, deploying, resolving something flagged as an anomaly — routes through an explicit human approval step before it executes. No generic bypass path.
- Audit trail
- An append-only log of every action that matters, so automated never becomes unaccountable.
Connections
- Human directs intent; Claude / Claude Code reasons and drafts the action.
- The drafted action is handed off to n8n, the deterministic orchestration layer.
- n8n enforces the approval gate as a concrete workflow step, not something the agent can skip.
- The human approves or denies the gated action before anything executes.
- Once approved, the action executes and n8n writes an append-only entry to the audit trail.
- The human (or the agent) reviews the audit trail, including anything flagged as drift.
Current focus
- Senior Big Data Administrator at Redsys
- AI agent & automation systems, under Aidtake Labs
Experience
From production data platforms to agent systems
Seven years administering production data platforms in banking and payments, most recently as a Senior Big Data Administrator at Redsys — the day job underneath the systems on this site.
- 01IT support, 2017–2019Started in systems and database support for a Santander environment (Everis) — the production-operations habits everything after this builds on.
- 02Big Data administration, 2019–2022Administered and migrated Cloudera clusters across Orange Bank, Santander, and the Bankia–CaixaBank integration — high availability, Kafka, hybrid on-prem/AWS environments.
- 03Senior Big Data administration, 2022–presentAt Redsys: end-to-end ownership of a major-version platform migration, platform security (authentication, authorization, audit, certificates), and L3 production support for payment infrastructure.
- 04Independent systems practiceThe same discipline — approval gates, audit trails, safe failure — carried into my own AI agent and automation systems, shipped under Aidtake Labs.
How I think
Principles, not slogans
Four rules that show up in every system above, in plain language.
- 01
Approval gates on anything irreversible.
Publishing, deployment, and anything irreversible keeps an explicit human approval step. Automation removes repetitive work, not the judgment call.
- 02
Audit trails, not just automation.
Every action that matters is logged in a way that can't be quietly edited after the fact, so "automated" never becomes "unaccountable."
- 03
Claude Code for judgment, n8n for deterministic glue.
The reasoning layer and the orchestration layer stay separate on purpose — a step a deterministic workflow enforces is a step an agent can't quietly skip.
- 04
Ship one correct system over ten demos.
I'd rather ship one system that's actually correct than ten that look good in a demo.
Knowledge
Two angles on the same systems
Same underlying work, written up for two different questions.
Orchestration
AI Automation & Agent Systems
The stack and the operational angle: Claude Code, n8n, and a data-infrastructure mindset carried over from Big Data / Cloudera work — what a business or consulting lead needs to know.
Read AI AutomationArchitectureAI Agent Architecture
The governance layer underneath: approval gates as a design pattern, identity and scope checks, immutable artifacts, and audit trails — for anyone evaluating whether an agent system is safe to give real access to.
Read AI Agent ArchitectureContact
Get in touch
For questions about any of these projects, including Traffic OS and its use of YouTube API Services, see the contact page.