Generative AI that answers from your own knowledge

A focused pilot, such as a support assistant grounded in one knowledge base, typically takes four to eight weeks from kickoff to a production release: one to two weeks of discovery and data preparation, two to four weeks of build and evaluation, and one to two weeks of integration and hardening. Larger programmes with fine-tuning, multiple data sources, or several integrations run twelve weeks or more. We agree the scope and dates before work starts and report progress weekly.

Three things: the documents or data the assistant should draw on (PDFs, help-centre articles, policies, product data, ticket history), a sample of the real questions people ask, and a stakeholder who can judge whether an answer is right. Access to the systems we will integrate with, such as your website, Slack, Teams, or CRM, comes later in the project. You do not need in-house AI expertise; we bring that.

We build on leading hosted models from OpenAI, Anthropic, and Google, and on open-weight models such as Llama, Mistral, and Qwen when data residency, cost, or fine-tuning calls for it. Around the model we use vector databases such as pgvector, Pinecone, and Weaviate, orchestration libraries such as LangChain and LlamaIndex, and cloud platforms including AWS Bedrock, Azure OpenAI, and Google Vertex AI, with Canadian regions where residency matters. We choose per project, based on a bake-off on your actual questions, and we are not tied to any vendor.

Before launch we build an evaluation set from real questions with agreed correct answers, and every change to prompts, retrieval, or model is scored against it for accuracy, groundedness, and refusal rate. After launch you get a dashboard and a monthly report covering answer quality, resolution or task-completion rate, user ratings, latency, and cost per conversation, plus a review of flagged conversations so the system keeps improving rather than drifting.

You do. Model-provider and cloud accounts are opened in your organisation's name with your billing, and we work as authorised users. All prompts, retrieval code, evaluation sets, and fine-tuned weights are delivered to your repository under your ownership, with the documentation to run them. Your data is used only to build your system; it is never used to train models for anyone else, and we configure each provider so your inputs are excluded from their training.

As an extension of it. We run a shared channel and a weekly review, pair with your developers on integration, and involve your subject-matter experts in scoring answers, since they know what right looks like. If you have an IT or security team, we bring them in early on data handling and access. At handover we document the system and train your team to maintain prompts, add documents, and read the monitoring, so you are not dependent on us unless you want to be.

It depends on the number of data sources, the integrations involved, whether fine-tuning is needed, and the volume of conversations you expect, so we scope each project individually rather than publishing a single rate. Ongoing model and hosting usage is billed by the providers directly to your own accounts, and we design for cost efficiency from the first prompt. Request a quote with a short description of your use case and we will reply within one business day with a proposed scope, timeline, and quote.