AI & machine learning
Our team builds with frontier models and runs open-weight models on local GPUs where data must stay in-house. We train and fine-tune models, stand up inference servers, and build retrieval and agent systems around them.
Our team builds with frontier models and runs open-weight models on local GPUs where data must stay in-house. We have trained and fine-tuned models, set up routing so each request goes to the cheapest model that can do the job safely.
Delivery pairs the engineering with governance: which data may reach which model, who approves what, and logs you can show to a client or a regulator. "Local AI vs Frontier APIs" compares cost, privacy and control, and the law firm case describes AI built into every layer of a custom platform.
What we do
- Local GPU servers for private inference
- Fine-tuning and training of language models
- Retrieval systems over company documents
- AI agents that call business tools safely
- Model Context Protocol servers for internal systems
- Comparing frontier and local models for a task
- Evaluation sets to measure answer quality
- Access controls and logging around AI use
We also run private research on open-weight models, including community-modified variants, in isolated environments. Clients get a clear view of what runs where, which data it can see and what it costs.
This group is the technical base of the Super Intelligence practice. It connects to Cyber Automation, where AI reads email, voice and documents, and to Cybersecurity, where we review the risks of connecting models to business data.
Leading platforms we work with
- Qwen
- DeepSeek
- Mistral AI
- NVIDIA
- Anthropic Claude
- OpenAI
- Google Gemini
- Meta Llama
- Zhipu AI GLM
- Hugging Face
Platforms and practices
- Local GPU servers: NVIDIA GPUs & CUDA
- Open-weight & abliterated models (private research)
- LLM training & fine-tuning from scratch
- PyTorch
- vLLM
- llama.cpp
- Ollama
- Retrieval-augmented generation (RAG)
- AI agents & tool use
- Model Context Protocol (MCP)
References
Sources last checked .
Product names are trademarks of their owners and are listed to describe our experience; they do not imply endorsement or partnership.
How we use AI & machine learning for clients
Frequently asked questions
Direct answers to the questions we hear most. Something else on your mind?
Send us a messageDo you train or fine-tune models?
Yes. The owner reports our team has trained LLMs from scratch and fine-tuned models. The right approach depends on your data and goal, and we recommend the simplest option that works.
Which models do you work with?
Open-weight families such as Qwen, DeepSeek and Mistral, and frontier APIs where appropriate. Our article on local AI versus frontier APIs compares them.
When should AI run locally instead of through a frontier API?
When data cannot leave your control, when volume makes API cost high, or when you need predictable latency offline. Frontier APIs are generally more capable and cheaper to start. We test your actual task on both before recommending, because the gap varies by task.
What is RAG and why use it?
Retrieval-augmented generation looks up relevant passages from your documents and gives them to the model with the question, so the answer is grounded in your content and can cite its sources. It avoids retraining the model whenever documents change. Quality depends on how the documents are split, indexed and permissioned.
How do you keep AI agents from doing harm?
Give each agent only the tools and data its job needs, require human approval for actions that move money or change systems, and log every tool call. Treat model output and retrieved text as untrusted input, since it can contain injected instructions. Limits and a kill switch matter more than prompts.
What hardware do local models need?
The main constraint is GPU memory, since the weights must fit; quantization reduces size at some cost in quality. Smaller models run on a single workstation GPU or a recent Mac, while the largest need several data-center GPUs. We size it from the model and the number of simultaneous users.