Agentic AI architecture
Design scalable, testable and evaluable multi-agent platforms using framework-agnostic patterns that can be adapted across industries, models and use cases.
I make ambitious AI ideas real — and give them a little soul.
Part engineer, part product thinker, permanently curious. I like giving AI a useful job, a clear personality and the infrastructure it needs to survive outside a demo. I lead projects, shape architectures, write Python and care about the small product decisions that make a complex system suddenly feel obvious.
JJ here
Currently fascinated by agents, intelligence and complex systems
AGENTS WITH A JOB ✦ SYSTEMS WITH SOUL ✦ IDEAS THAT SHIP ✦ CURIOUS BY DEFAULT ✦
The kind of problem I can’t leave alone
Can we make AI useful when the answer actually matters?
My work has converged on autonomous agents, information retrieval, scalable infrastructure and human-in-the-loop operations. I want to keep building systems that help people investigate complex evidence, coordinate specialised tools and make better decisions when the answer genuinely matters.
Engineering agents from prototype to productionProduction AI infrastructure
Architect and operate provider-agnostic, multi-cloud AI platforms with repeatable delivery, observability, evaluation and lifecycle management built in from the start.
Information systems at scale
Design distributed information systems that ingest, connect and serve large volumes of heterogeneous data through reliable processing, retrieval and real-time interfaces.
End-to-end product ownership
Lead projects from product thinking and problem framing through proof of concept, MVP, production launch and continuous evolution — connecting user needs, architecture and delivery.
Chapter 01
My in-house career
Experience
My in-house career
From hands-on builder to AI platform lead.
I’ve worked across data analysis, data engineering, frontend, backend, cloud and AI engineering. That breadth made me a technical Swiss Army knife — able to understand the whole system and step into whatever role a project needs.
2025 — Present
Insud Pharma · AI Labs
Tech Lead AI Engineer / Architect
Created and lead Axon, an agent industrialisation initiative spanning reusable templates, CI/CD, Langfuse observability and evaluation. The platform supports more than ten healthcare assistants plus multi-agent systems for factory operations, business intelligence and pharmaceutical quality workflows.
Turned reusable agent patterns, delivery workflows and engineering standards into a shared platform that helps teams move from an AI idea to an observable, evaluated and maintainable product.
- Python
- LangGraph
- Deep agents
- Langfuse
- Azure
- AWS
- Elasticsearch
- Qdrant
- OpenAI
- Gemini
- Terraform
2023 — 2025
Telefónica Global Solutions
Lead AI Engineer / Architect
Led a team designing a scalable end-to-end generative AI platform and the company’s first autonomous AI telecom operator, HELI. Architected multi-agent, cloud-native services and a real-time assistant spanning chat, tasks, documents and call-centre workflows.
Earned IT’s trust to manage my team’s AWS infrastructure, from permissions and IAM roles to the deployment of pipelines and cloud resources.
- Python
- LangGraph
- RAG
- AWS
- Bedrock
- SageMaker
- MLOps
- EKS
- Kubernetes
- Elasticsearch
- Spark
- Hadoop
- EMR
- WebSockets
2019 — 2023
Telefónica Global Solutions
Full-stack Software Engineer
Created backend services, AWS architectures, Python libraries, CI/CD pipelines, distributed ingestion processes and EKS deployments. Developed Kibana interfaces for NLP chat, logs, notifications and operational workflows, working alongside Spark, Hadoop and EMR-based data platforms.
Learned the data and application lifecycle from every angle — data analysis, data engineering, AI engineering, frontend, backend and cloud — giving me a genuinely global view of how systems fit together.
- AWS
- Python
- Spark
- Hadoop
- EMR
- SQS/SNS
- EKS
- Linux
- Kibana
- React
- CI/CD
Chapter 02
Company, freelance and personal work
Projects
Company work · Insud Pharma AI Labs
Featured projects at Insud Pharma.
Professional projects developed as part of my role at Insud Pharma AI Labs.
01
Enterprise agent platform · 2025 — Present
Axon
The factory behind a growing ecosystem of production agents.
An AI-agnostic initiative that turns proven agent patterns into accessible templates for technical and non-technical teams, while automating deployment, observability and evaluation.
- My contribution
- Created the initiative and lead its product direction, platform architecture and hands-on engineering across the complete agent lifecycle.
- Outcome
- Enabled 10+ internal and public healthcare assistants plus advanced multi-agent systems for industrial operations, executive intelligence and pharmaceutical quality.
Inside the work · Insud Pharma AI Labs
Axon is not one agent. It is the system that makes many agents possible.
I created Axon as an ambitious initiative to democratise agentic AI across the organisation. The goal is to make reliable agents accessible to technical and non-technical teams through simple, reusable templates — without turning every new use case into a bespoke engineering project.
A prototype is only the beginning. Axon industrialises the full path from an idea to an observable, evaluated and maintainable production agent.
Compose
Reusable templates and opinionated patterns for chatbots, orchestrators, specialised agents and advanced deep-agent systems.
Deploy
Automated CI/CD and infrastructure patterns that make releases repeatable instead of dependent on individual projects.
Observe
End-to-end tracing and operational visibility, so teams can understand what an agent did and why it did it.
Evaluate
Langfuse-powered evaluations that turn quality, regressions and behaviour into something measurable and improvable.
The advanced pattern
An orchestrator with a team — not one overloaded agent pretending to know everything.
For complex workflows we built a LangGraph multi-agent architecture similar to today’s deep-agent pattern. A central orchestrator delegates to specialised agents, each with an independent context and a clear responsibility. This makes individual behaviours easier to evaluate, reduces context pollution and limits the risk of one agent attempting too much. The pattern is now part of Axon’s advanced-agent factory using the deep-agent standard.
- Independent context
- Explicit responsibility
- Focused evaluations
- Lower coordination risk
Axon / in production
What Axon has already made possible
01
10+ internal and public assistants
Healthcare chatbot factory
A repeatable family of assistants for healthcare brands and teams, built on shared patterns rather than isolated implementations.
- Bloom / BeBloomers — A public web and WhatsApp health assistant handling hundreds of sensitive requests each week.
- Slynd — A product-information assistant in QA, designed to answer questions from the website and ultimately from QR codes on medicine packaging.
- Exeltis Colombia — An internal assistant supporting medical professionals with product information.
02
Multi-agent factory operations
Industrial Copilot
An advanced agent connected to factory knowledge and operational state. It can inspect machines across production and delivery, troubleshoot why equipment stopped and investigate likely causes using the complete documentation context.
- Production agent — explores production lines, machines and SCADA telemetry including status, temperature, OEE, performance, quality and parts.
- Maintenance agent — uses RAG over technical factory documentation to investigate preventive and corrective maintenance tasks.
- Replacement agent — checks inventory, shipments and inter-factory logistics to locate parts and estimate availability and delivery.
03
Multi-agent research and executive reporting
Market Intelligence
A multi-agent intelligence system where a master orchestrator coordinates specialised agents across distinct data sources, combines their findings and turns the resulting evidence into analyses and executive reports.
- Snowflake agent — explores structured business data including sales, patents and related commercial information.
- Elastic agent — retrieves medicine information, technical data sheets and other indexed product knowledge.
- Browser agent — researches the open web when internal sources are not enough.
- The orchestrator reconciles results and produces data analysis, web reports, executive documents and presentations.
04
Complaints, deviations and CAPA
TrackWise Quality
A specialised quality agent connected to Salesforce/TrackWise data for investigating complaints and deviations throughout factory processes. It helps explore root causes, structure investigations and generate corrective and preventive action proposals.
- Read-only search and exploration across complaints and manufacturing deviations stored in Salesforce/TrackWise.
- Root-cause investigation grounded in quality-system evidence.
- CAPA proposals to resolve issues and reduce the chance of recurrence.
Company work · Telefónica Global Solutions
My work at Telefónica.
Professional work delivered during my career at Telefónica Global Solutions.
01
Autonomous telecom operations · 2023 — 2025
HELI
An autonomous AI operator built for real operational work.
A multi-agent generative AI platform integrating conversational assistance, tasks, notifications, documents and call-centre workflows within a scalable cloud-native architecture.
- My contribution
- Designed the product, led the engineering team, established the architecture and translated a broad operational vision into independently deliverable systems.
- Outcome
- Established Telefónica Global Solutions’ first autonomous AI telecom operator and a reusable foundation for AI-enabled operations and delivery.
My own projects
My personal projects.
Personal products and experiments where I own the concept, architecture, implementation and evolution.
01
Voice AI · Founder project · 2022 — Present
MIA Bot
The world’s first LLM integration for Alexa.
An early Alexa Skill product connecting a large language model to a voice-first interface, built when conversational generative AI was only beginning to reach mainstream users. What started as a prototype now runs in production as an actively maintained service that scales automatically with demand.
- My contribution
- Founder, product design, architecture, implementation, release and ongoing operation.
- Outcome
- Reached more than 20,000 inputs per week, over 2,000 active users and 10–30 new users per day.
02
Autonomous product factory · 2025 — Present
Aiscraft
A company operating system designed around agents.
A personal R&D platform where autonomous agents create, deploy and monitor software products using reusable application libraries and infrastructure modules.
- My contribution
- Product concept, system architecture, agent workflows, cloud infrastructure and the applications built on top.
- Outcome
- A working environment for rapidly exploring production AI patterns through products including Travel Radar and Pokémon Team Builder.
03
Legal AI prototype · 2024
Estado.ai
Making complex Spanish legal information easier to understand.
01 / 02
An AI-powered legal information platform that ingests open Spanish public data and makes laws, notices and regulations explorable through search, dashboards and generative AI.
- My contribution
- Created together with Alejandro Sánchez Losa, Elasticsearch Solutions Architect. I contributed to the concept, product design, BOE ingestion, Elasticsearch architecture, retrieval workflows and full-stack prototype.
- Outcome
- Turned a large public-data corpus into a searchable and observable system, exploring how retrieval and language models can make a high-friction information domain more approachable.
04
Personal engineering method · Agent evaluation · Ongoing R&D
Autonomous Agent Optimisation
A way of evaluating agents that can improve what it measures.
A personal evaluation approach I use to make agent experiments comparable and actionable. It is deliberately simple, flexible and independent of any specific agent framework or harness: define the goal, measure behaviour against explicit criteria and use the evidence to guide the next improvement.
- My contribution
- Designed the method from scratch, including experiment definition, requirement verification, measurable technical and product outcomes, and an optional improvement loop that can propose changes to prompts, configuration, models or examples.
- Outcome
- Provides a reusable way to evaluate any agent or experiment, understand regressions and evolve behaviour through evidence instead of intuition alone.
Freelance projects
Selected freelance projects.
Selected independent projects created for students and private clients, from product thinking to working software.
01
E-commerce · Final degree project · Freelance
Miss Margot
An online shop where finding a product could begin with an image.
A complete computer science final degree project built around an e-commerce experience with visual product discovery. Users could submit an image and use Google Cloud Vision capabilities to help find related products in the catalogue.
- My contribution
- Product definition, architecture, visual-search integration and full-stack implementation for the client’s final degree project.
- Outcome
- Combined a conventional commerce journey with an accessible demonstration of applied computer vision and image-based search.
02
Web product · CV automation · Freelance
CVSite
Turning structured career information into a ready-to-use CV.
A web application created to make CV creation faster and more consistent by generating the document automatically from the information provided by the user.
- My contribution
- Product design, information model, automated document-generation flow and full-stack implementation.
- Outcome
- Transformed a repetitive formatting task into a guided web workflow with a reusable, automatically generated result.
03
Trading automation · Data engineering · Freelance
Multi-broker Trading Bot
From fragmented market information to automated, traceable execution.
A trading automation system that collected information from multiple websites, normalised the resulting data, calculated quantitative metrics and signals, and executed orders through integrations with several brokers.
- My contribution
- Designed and implemented the complete workflow: resilient web scraping, data processing, metric calculation, decision rules, broker adapters, order execution and operational safeguards.
- Outcome
- Unified research, analysis and execution in one repeatable pipeline while keeping broker-specific behaviour isolated behind reusable integrations.
My toolbox
01AI
Agent & ML systems
- LangChain & LangGraph
- Custom, deep & ReAct agents
- RAG & embeddings
- Model evaluation: Langfuse, LangSmith
- MLOps
- SageMaker
- Bedrock
- Multi-model providers: OpenAI, Gemini, Claude…
02{ }
Software & data
- Python
- TypeScript, JavaScript
- Node.js
- React, Next.js
- Embedding stores: Elasticsearch, Qdrant, Chroma…
03☁
Cloud & operations
- Multicloud: AWS, Azure, GCP
- EKS / Kubernetes
- Serverless
- Terraform
- Docker
- CI/CD
- Observability
- Distributed systems
Chapter 03
Studies, certifications and foundations
Education
Proof I occasionally study on purpose
Engineering foundations, with a lifelong side quest.
Universidad Politécnica de Madrid
BEng in Telecommunication Technologies and Services
Specialisation in telematics. Final degree project graded 10/10.