Local AI
On-premise hardware plus standardised inference interfaces. Private AI that never has to leave your building.
Modern AI is real leverage, but the standard way of getting it means sending your material through someone else’s API, on their roadmap, under their retention policy. For a lot of teams that’s fine. For some it isn’t. We install and implement on-premise AI hardware with a standardised inference interface on top, so the same client code can talk to your box that would have talked to a hosted vendor, and the data never leaves the building.
Hardware
- Specification. Model sizes, concurrency, throughput and latency turned into an actual bill of materials: workstation, rack, edge box, or a mix. GPU class chosen for the models you’ll actually run, not the ones on the datasheet.
- Procurement and install. We source, rack and cable it, or work with your infrastructure people who already do. Networking, storage, power, cooling, remote management. The physical layer sorted, not assumed.
- Air gap or perimeter. We agree with you where the boundary sits: fully air-gapped, VLAN-isolated, or reachable only from your internal network, and we build so the system enforces that at runtime rather than in a policy PDF.
Inference stack
- Serving. vLLM, Ollama, llama.cpp or TGI for language models, plus local embeddings and (optionally) speech and vision. Model chosen for the job, not the hype. Config committed, reproducible from a fresh box.
- Standardised interface. An OpenAI-compatible HTTP endpoint on your box. Existing SDKs, IDE extensions, agent frameworks and internal tools point at it by URL and stop caring what runs behind it. Swap models without touching client code.
- Retrieval. Local vector store, hybrid search, and an evaluation harness, so retrieval quality is a number you can watch instead of a feeling.
Handover
- Operations. Runbooks written for the on-call person you actually have, not the SRE team you don’t. Backup, upgrade, incident, capacity.
- Evaluation. A small suite of your own prompts that runs on every model change, so “we upgraded and it feels different” becomes a measurable claim.
- Cadence. The open-weights space moves fast. We agree on a check-in rhythm so the setup doesn’t quietly age past useful.