I design and deploy private AI infrastructure that keeps your data under your control. No vendor lock-in, no API rate limits, no telemetry. Everything runs on hardware you own.

My approach is infrastructure-first: Dockerized stacks, modular compose files, GPU-accelerated inference, and security baked in from day one. Every deployment is air-gap-ready by default.

Need a private AI setup? Send an email to consulting@blob42.xyz and we can discuss your use case.

Schedule a call now

Infrastructure Setup

I provision and configure the full stack from bare metal to application layer. Every component is containerized, version-controlled, and designed for reproducibility.

• GPU Server Provisioning: Ubuntu-based instances with NVIDIA CUDA/AMD ROCm, configured on AWS, Runpod.io, or your own hardware. I handle bare metal installation (OS, Hypervisor, Docker/Podman Host), CUDA toolkit setup, and GPU resource allocation.
• Docker Compose Architecture: Modular, flat .yml files with one service per logical unit (LLM engine, UI, RAG, agents). No monolithic compose files — each component is independently startable and replaceable. (Also works with Podman)
• Reverse Proxy & Networking: Caddy reverse proxy with automatic TLS (internal PKI or Let’s Encrypt), name-based virtual hosting, and granular access control. Internal Docker networks isolate services from external exposure.
• Model Storage & Caching: Persistent volume strategy for GGUF models, KV caches, and HuggingFace download caches. Optimized for fast model loading and low-disk-footprint operation.
• CI/CD for Models: Automated pipelines for model updates, quantization, and benchmarking. Built from local source with custom CUDA architectures (Turing through Hopper).

ELI5: Think of this as building the physical foundation and plumbing. Instead of renting a small apartment in the cloud, I build you a secure, custom-made fortress on your own land. You own the keys, the walls, and the power.

Model Selection & Inference

I help you pick the right model for your workload and deploy it with the optimal inference engine. I run multiple inference backends in production and can recommend the best fit for your constraints.

• Inference Engines: llama.cpp (direct, closest-to-metal), LocalAI (full OpenAI-compatible API, multi-model), vLLM (high-throughput serving), Ollama (simplicity-first). I contribute directly to the LocalAI codebase.
• Model Hot-Swapping: llama-swap proxy for automatic loading and swapping between models without downtime. Config-driven, supports multiple concurrent model servers.
• Model Families: Gemma, Qwen3, Llama, Mistral, and others in GGUF format. I benchmark models on your hardware to find the best trade-off between quality, speed, and VRAM usage.
• Quantization: Model quantization (Q4, Q5, Q6, Q8) to fit larger models on smaller GPUs. Custom quantization pipelines using llama.cpp tools.
• Multi-GPU & Multi-Arch: Builds targeting specific GPU architectures (CUDA 86 for RTX 20/30/40 series, 89 for Hopper). Static binary builds for portable deployment.

ELI5: This is choosing the “brain” of your AI. Some brains are huge and smart but slow; others are small and fast. I find the perfect balance so your AI is smart enough for the job but doesn’t lag.

RAG Pipelines

Retrieval-Augmented Generation (RAG) is the bridge between your private documents and the LLM. I build end-to-end RAG pipelines that eliminate hallucinations by grounding AI responses in your actual data.

• Vector Database: ChromaDB for self-hosted, lightweight vector storage. Document ingestion, embedding, and similarity search out of the box.
• Embedding & Reranking: Model selection for embeddings and reranking. I match the embedding model to the nature of your data (technical docs, legal contracts, etc.) to ensure high retrieval accuracy.
• Chunking Strategy: Domain-specific tuning of document chunking size and overlap. I optimize how data is “sliced” so the AI gets the exact context it needs without wasting tokens.
• UI Integration: Open-WebUI with built-in RAG document upload. Users can ingest PDFs, text files, and web pages directly through the interface for immediate grounding.

ELI5: This is giving the AI a “library” of your company’s documents. Instead of the AI guessing or making things up, it looks up the answer in your books first, then summarizes it for you.

Agentic Systems

Beyond simple chat, I deploy autonomous agent workflows that transform the AI from a consultant into a worker that can execute tasks across your infrastructure.

• Agent Frameworks: LocalAGI for local-first orchestration, N8N for self-hosted workflow automation with built-in connectors (email, APIs, databases), and LangChain for complex custom logic.
• Custom MCP Servers: Model Context Protocol servers tailored to your internal tools, allowing the AI to read your databases or trigger internal scripts securely.
• Tool Development: Custom pipes and valves for Open-WebUI, actions for LocalAGI, and connectors for N8N. I build the “hands” the AI needs to interact with your business.

ELI5: This turns the AI from a “chatbot” into a “digital employee.” Instead of just telling you how to do something, the AI can actually do it, like sending an email or updating a database.

Diffusion Models (Image Generation)

For visual workflows, I deploy and containerize diffusion model stacks that allow you to generate high-fidelity imagery and assets privately.

• ComfyUI: Node-based Stable Diffusion interface with GPU acceleration in Docker. I provide optimized Dockerfiles with NVIDIA CUDA support for professional-grade image generation.
• Stable Diffusion WebUI: Alternative UI with multiple build variants tailored to your specific hardware constraints.
• Model Management: LoRA, checkpoint, and embedding organization across persistent volumes for consistent brand and style control.

ELI5: This is your own private AI artist. You can create images, logos, or concept art without sending your prompts or results to a third-party company.

Security & Privacy

Security is the core of my methodology. Every deployment is designed with a privacy-first, air-gap-ready mindset to ensure your intellectual property never leaves your perimeter.

• Network Isolation: Docker networks segment traffic by function. Internal services communicate over private networks, ensuring that a compromise in one layer doesn’t expose the entire stack.
• Authentication: SSO integration via LDAP (LLDAP), Anubis authentication gateway, and robust API key management for programmatic access.
• mTLS & Internal PKI: Mutual TLS between services using self-signed certificate authorities. This ensures that only authorized services can talk to each other.
• Secret Management: Docker secrets for sensitive material and encrypted configuration via git-secret, ensuring that passwords and keys are never stored in plain text.
• Air-Gap Deployment: Full stacks designed to operate on isolated intranets with zero external connectivity.
• Hardening: SSH key-only access, strict firewall rules, and minimal port exposure.

ELI5: This is the vault. I make sure your data is locked down so tightly that not even a single byte leaks out of your building.

Integration & Tooling

I ensure your AI stack integrates seamlessly with your existing software ecosystem and developer workflows.

• OpenAI-Compatible API: All backends (LocalAI, llama.cpp, vLLM) expose OpenAI-compatible endpoints. This allows you to swap out expensive cloud APIs for your local stack with zero code changes.
• LLM Debugging Proxy: lightweight vendor free MITM proxy for inspecting LLM calls, context windows, and tool invocations. Essential for auditing AI behavior and optimizing prompts.

ELI5: This is the glue. I make sure the AI can “talk” to your existing software and tools, so you don’t have to change how you work to use AI.

Engagement Models

I offer flexible engagement models depending on your needs:

• POC / Proof of Concept: Rapid deployment of a working private AI stack in 2-4 days. Validated with a live demo before full commitment.
• Full Deployment: End-to-end design, provisioning, and handoff of production-ready infrastructure with full documentation.
• Retainer: Ongoing support, model updates, and optimization. 30-day+ retainers available for continuous collaboration.
• Custom Development: Tailored tools, integrations, and agent workflows specific to your use case.

The Legal Stuff

All services are provided by me as a registered autónomo in Spain, with ROI registration for intra-EU VAT.

Invoices are generally issued in EURO. I am available for engagements up to 4 days per week.

VAT: ES9065609N