Pol Fernández · AI Systems Engineer
AI systems that keep working when nobody is watching.
I design and build agentic systems for real operations, grounded in controlled data, connected to existing tools and built with end-to-end observability.
In 5+ years of building software, I have taken 20 systems into production.
From model to system
The model is one component. The system is the work.
Production starts where the demo ends. An AI system must know what it may access, which tool it should use, what changed and how to recover when a dependency fails.
I design those constraints into the architecture from the first line.
- 01
Orchestrate
Specialised agents with explicit responsibilities, shared state and deterministic handoffs.
- 02
Ground
Controlled document sets, source citations and retrieval paths that can be inspected.
- 03
Integrate
Existing APIs, repositories, queues and test infrastructure connected without replacing what already works.
- 04
Operate
Cost, latency, decisions and failure paths traced on every run.
Technology follows the constraint.
The stack spans models, orchestration, data, delivery and infrastructure.
- Python
- TypeScript
- PostgreSQL
- Docker
- Kubernetes
- Linux
- Git
- GitHub Actions
- Azure
- Cloudflare
- Playwright
- n8n
- Claude
- LangChain
- Ollama
- Proxmox
Flagship system
Pharma and laboratory robotics
Seven agents. Fifty two tools. One accountable test system.
A legacy test suite covered instrument software used in the handling of patient samples. Its replacement coordinates specialised agents across the tools engineers already use, with one supervisor controlling the run.
Every run records cost, duration and the changes produced. Each agent is assessed against measured contribution, not a staged demonstration.
7specialised agents
52+connected tools
3model backends
Control plane
Supervisor
One toolbox. One execution context. A trace for every run.
- Claude
- Anthropic API
- Ollama
Specialised execution paths
- UC-01 Order generator
- UC-02 Log management
- UC-03 Deck visualiser
- UC-04 Protocol parser
- UC-05 Pull request reviewer
- UC-06 Knowledge consolidation
- UC-07 Test orchestration
Selected work
Different domains. The same production standard.
The architecture changes with the constraint. The engineering standard does not.
- 01
Defence and optronics
Retrieval over classified documentation
Air-gapped deployment, source citation on every answer and a complete audit trail of who queried what.
- 02
Health and wellness
CRM with access control enforced in the database
Row-level security in PostgreSQL, with a policy test per role and field-level encryption.
- 03
Pharma and laboratory robotics
Quality gates in the release pipeline
Containerised runners and a virtual machine snapshot per execution, so every failure is reproducible.
- 04
Research and prototyping
3D scanning and robotic arm control
Geometry acquisition, motion planning and a simulator for collision checking before execution.
- 05
Internal systems
Scheduled agents with cost reporting
Daily review of incoming signals and automated transfer of time records between two operational systems.
Built and operated
I build the system. I stay close to how it runs.
I am Pol Fernández, an AI and software engineer based in Au on Lake Zurich. I work in German, English and Spanish, from architecture through deployment and operation.
My own production-equivalent environment runs on dedicated hardware under my administration. Identity, certificates, monitoring and restore procedures are live systems. I test failure and recovery there before applying changes to a client environment.
- Identity
- Single sign-on across every service
- Exposure
- No service exposed by default
- Alerting
- Before user impact
- Recovery
- Six backup layers, verified by restore
Contact
Have a system that needs to survive real use?
Send me the problem, the constraints and what already exists. I reply within one working day, including a direct answer if it is outside what I can usefully contribute to.