Self-hosted LLMs
Open-weight models served on your own GPUs or Apple Silicon, behind an OpenAI-compatible API your team already knows how to use.
Local AI & ML engineering consultancy
AI that runs on your hardware, not someone else's.
I help companies deploy private, self-hosted AI — open-weight LLMs on your own servers, on-device models inside your apps, and ML pipelines that keep sensitive data in-house.
Built with
- vLLM
- llama.cpp
- Ollama
- MLX
- Core ML
- PyTorch
- ONNX
- YOLO
- Docker
Services
Everything between "we should use AI" and a system in production.
Hands-on engineering, not just strategy. I design, build and deploy — then make sure your team can run it.
On-device AI for apps
Quantized models running directly on iPhone, iPad, Mac and Android — summarization, chat and vision with no server round-trip.
Private RAG & search
Question-answering and semantic search over internal documents, mail and scans — indexed, embedded and served inside your network.
Agents on open models
Tool-using agents that run on local models, with routing that escalates to a frontier API only when a task genuinely needs it.
Computer vision pipelines
Detection, classification and tagging for images and video, from dataset labelling through training to batch or real-time inference.
Hardware & cost sizing
Which model, which quantization, which box. Benchmarks on your workload so you buy the hardware you need — and not more.
Why local
Frontier APIs are great. Sometimes they're the wrong tool.
Open-weight models now handle a large share of real business workloads. When privacy, cost or latency matter, running them yourself is often the better trade — and I'll tell you honestly when it isn't.
01
Your data stays put
Prompts, documents and outputs never leave infrastructure you control — simpler GDPR and client-confidentiality conversations.
02
Predictable cost
A fixed hardware budget instead of per-token bills that grow with every new user and every longer context.
03
Low latency, works offline
Inference next to the user or on the device itself. No network dependency, no third-party outage taking you down.
04
No lock-in
Open-weight models and standard APIs. Swap models as better ones ship, without rewriting your product.
Approach
A short path from idea to evidence.
Every engagement is built around measurable results on your own data, early.
- 01
Assess
We map the use case, data sensitivity, latency targets and budget, and decide whether local is the right call at all.
- 02
Prototype
A working proof of concept on your data, with measured quality, speed and cost — not a slide deck.
- 03
Deploy
Production serving, monitoring and evaluation set up on your hardware, cloud account or devices.
- 04
Hand over
Documentation, runbooks and a walkthrough so your team owns and operates it from day one.
Engagements
Pick the depth that fits.
Fixed-scope where possible, so you know what you're getting. Ongoing retainers available after delivery.
Advisory
A focused session to pressure-test a plan.
- Architecture & model selection review
- Hardware and hosting recommendation
- Written summary with next steps
Proof of concept
Two to four weeks to prove it works on your data.
- End-to-end prototype
- Quality, latency & cost benchmarks
- Go / no-go recommendation
Build & deploy
From prototype to something people rely on.
- Production deployment
- Evaluation & monitoring
- Team handover and support
About
Hi, I'm Alexandru Tudorica.
I'm a software engineer who has spent years building systems end to end — from backend services and infrastructure to native mobile apps. These days I focus on making modern AI practical to run privately: serving open models on workstations and servers, shipping on-device inference in iOS apps, and wiring up the pipelines around them.
You work directly with me, from the first call to the handover. No account managers, no junior hand-offs.
Writing
All postsFrom the blog
bicameral: Claude does the thinking, a local model does the typing
An open-source tool that lets Claude Code hand mechanical work to an open model on your own machine. Free tokens, files that stay local, and Claude still checks the result.
Read →When does it make sense to run LLMs on your own hardware?
A practical checklist for deciding between a hosted API and self-hosted open-weight models — and the hidden costs on both sides.
Read →Have a use case in mind?
Send a few lines about what you're trying to do, what data is involved and any constraints. I'll reply within two business days, usually with a few questions and a suggested next step.