What millAGI is
millAGI is my consulting practice. I am Javier Ladrón de Guevara: 30 years in systems, infrastructure and security, and now deep into language models.
The name is a wink. Nobody knows when AGI arrives. What I do know is how to build the AI a company can use today: on its own machines or outside, measured on its own data, and without slide-deck promises.
Everything offered here already works, and each service comes with real examples. If I cannot show it, I do not sell it.
Services
Private AI
Language models on your own machines, with no data sent to third parties. I pick the hardware by measuring before you buy, set up the inference engine, quantize the models and put a single entry point in front, with your company login, per-team keys and spending caps.
Examples: depending on each case, I have deployed open models such as DeepSeek, GLM or Qwen for up to 120 users (30 at the same time), on hardware that goes from AMD Strix Halo to MI350 or H200, with R9700 in between.
The best model, every week
The best model for a task changes almost every week. I help you always use the one that fits: I measure it on your real tasks, not on public leaderboards, and measure again when something new comes out. Speed, quality, cost per task and behaviour under load. I also review subscriptions: what you pay, what you use and what you can drop.
Examples: Model Speed Arena measures speed and cost across providers; Yardstick measures open-model quality with each benchmark's official protocol, its error margin and the energy spent.
Process automation
Take a repetitive process to production with language models: ingestion, summary, classification, human review and publishing. I help decide which part needs a model, which part is a plain script and where a human must stay in the loop. It ships with metrics, retries and alerts.
Examples: self-maintaining code and automatic content cataloguing, following industry best practices. One of the pipelines went from 37 model calls per task to less than one, with the same result.
Custom models
I train a small model for one concrete, repeated task. It is cheaper and faster than a large model, and it runs on modest hardware. I prepare the data, distil from a large model, fine-tune, quantize and measure how much quality is lost before deploying.
Examples: 2B and 4B models that summarise and classify the news on ojeo, and a 2B model that rewrites text in a given voice.
Coding with agents
I get your development team working with Claude Code, opencode and similar tools, with control: permissions, per-repository rules, code review with several models and automatic dependency maintenance. I standardise the process: what the agents do becomes part of the software development lifecycle, with the same reviews and security checks as everything else. With hands-on workshops.
Examples: a system that reviews and merges dependency updates across 21 repositories. Workshops delivered.
AI security
Two fronts, on top of a classic offensive security background.
Privacy. Agents open new doors. I look for secrets and personal data that end up in the model context, uncontrolled internet access, excess permissions and instructions hidden in the data the agent reads.
Model behaviour. Test suites that measure the model's biases and whether it behaves as expected in the real use you will give it.
Examples: a proxy between the agents and the model provider that swaps secrets for a placeholder before they leave and measures which identifiers go out (IPs, host names, emails), the first step towards pseudonymising them. Years of audits, pentesting and training. OSCP and a master's degree in cybersecurity.
Infrastructure and DevOps
The base for everything above. I deploy these solutions at scale, on your own servers or in the cloud, and integrate them with the rest of your tech stack: identity, data, monitoring and continuous delivery. Kubernetes, Terraform, Ansible, CI/CD, AWS, Elasticsearch at scale and search over your own documents. Also as a fractional service: a few hours a month as your infrastructure team.
Examples: Kubernetes, CI/CD and Elasticsearch in production, with 30 years in the trade behind them.
How I work
| Format | What for | Price |
|---|---|---|
| Assessment | One concrete question: which model, which hardware, which process to automate. | Fixed |
| Implementation | One of the services above, running in production. | Per project |
| Follow-up | Regular review of models and providers: what came out and whether you should switch. | Monthly fee |
| Fractional | Hours per month of infrastructure or AI work. | Monthly fee |
| Workshop | Hands-on training for a team. | Per session |
What I do not do
- "AI transformation" strategy for executives with no technical case behind it.
- Training large or general-purpose models from scratch.
- Legal advice on the EU AI Act. That needs a lawyer; I can work with yours.
- Data science and classic machine learning.
- Custom web or mobile apps as the main product.
Public examples
Yardstick
Open-model quality measured with the official protocol, with its error margin, its quantization and its energy use.
Sectors
Today I work with clients in these sectors:
- Public sector
- Hardware and software security analysis
- Investigative journalism
- Product companies
Contact
Tell me what you want to solve. The first conversation is free and with no commitment.
Click a contact method to show it