Confidential AI

Most AI services are designed on the assumption that your data will travel to someone else's infrastructure. For professionals handling client confidences — a lawyer managing case files, a fiduciary holding financial records, an engineering team working on proprietary code — that assumption is not compatible with professional obligation.

Confidential AI is our end-to-end service for deploying AI where the data stays: on hardware you control, in a network you manage, with no external API calls and no data leaving your premises. We work from use case assessment through to running models in production — every step of the journey, for clients who cannot afford the alternative.

Use Case Assessment

Not every problem needs an AI model, and not every model needs the same architecture. Before writing a single line of code, we spend time understanding the actual workflow: what the professional does today, where the friction is, and what a good outcome looks like in practice.

The core question is architecture: does this problem call for a conversational AI — a system that responds to queries, surfaces information, and reasons over documents — or an AI agent — a system that takes actions autonomously, searching, drafting, filing, or executing workflows with minimal human instruction? Getting this wrong means building something technically sophisticated that does not match how the user actually works.

• Workflow analysis — understanding the daily practice: the documents, the queries, the decisions that repeat
• Use case fit — determining whether the value comes from retrieval, reasoning, generation, or automation
• Agent vs conversational AI — choosing the right architecture based on the workflow, not the current trend
• Risk and compliance review — mapping the use case against professional confidentiality obligations, data residency requirements, and applicable regulations

Model Selection and RAG Design

Local AI means running the model yourself — which means choosing it yourself. We evaluate open-weight models — Mistral, Llama, Qwen, Phi, and others — against the specific domain, language, and performance requirements of each engagement. A model well-suited to legal French is not the same as one suited to German fiduciary reporting or Swiss code documentation.

For most professional use cases, the AI's value comes from its ability to reason over the client's own documents — case files, contracts, correspondence, codebases — not just general knowledge. This requires a RAG pipeline — Retrieval-Augmented Generation — which retrieves relevant context from a private document store and provides it to the model at inference time, keeping every document on-premises.

• Model evaluation — benchmarking candidate models on domain-representative tasks before committing to a deployment
• RAG pipeline design — document ingestion, chunking strategy, embedding model selection, vector store design
• Context window management — handling long documents, multi-document reasoning, and citation tracking
• Prompt engineering — system prompts, few-shot examples, and instruction formats tuned to the professional domain

Application Development

The model and the pipeline are not the product. The product is the application that a lawyer, fiduciary, or developer actually uses — one that fits naturally into how they already work, surfaces results in a form they can act on, and fails gracefully when the model's confidence is low.

• Conversational interfaces — secure web applications for querying documents, drafting text, and reasoning over case materials
• Agent workflows — automated pipelines for document classification, deadline extraction, code review, or research summarisation
• Document processing — ingestion pipelines for PDF, Word, email, and structured data; OCR where required for scanned materials
• Integration — connecting to existing practice management systems, document stores, or code repositories without routing data through external services

Secure Deployment

A local AI deployment that is reachable from the internet, or that stores conversation history in plaintext, does not deliver on its security promise. We deploy on infrastructure hardened from the first day, designed for professional confidentiality obligations.

• Air-gapped and on-premise options — models running on hardware with no outbound internet requirement; deployment models that satisfy strict data residency requirements
• Model serving — Ollama for single-machine deployment; vLLM for multi-user throughput; LiteLLM for unified API access across models
• Authentication and authorisation — every endpoint behind Authentik SSO with MFA; role-based access to different models and document collections
• Encryption at rest and in transit — encrypted storage for model weights, document indexes, and conversation logs; TLS everywhere, internal CA for service-to-service trust

Hosting and Operations

Deploying a model is not the same as running it sustainably. We stay involved after launch — monitoring inference latency and throughput, managing model updates as better versions become available, and ensuring the system performs correctly as document volumes and user counts grow.

• Hardware sizing — GPU selection for inference throughput requirements; CPU-only deployment where GPU is not available or necessary
• Monitoring — latency, throughput, error rates, and token usage tracked continuously; alerting before performance degrades
• Model updates — evaluating new model releases against the current deployment, running comparative assessments, deploying updates with no service interruption
• Ongoing support — incident response, prompt refinement as edge cases emerge, and RAG pipeline tuning as the document corpus evolves

Who this is for

Legal practices, fiduciary and accounting firms, and software consultancies that handle client-confidential material and cannot send that material to third-party AI services. Organisations subject to data residency or professional secrecy obligations. Teams who have evaluated cloud AI tools and concluded that the data risk outweighs the convenience.

If you are currently doing work by hand that AI could assist with, but have not been able to adopt cloud AI for confidentiality reasons — that is exactly the situation this service is built for.