Support and sales
A chat assistant answers from your data and passes the next step to a person. An agent can take a longer path: look up an order, draft a reply, or call your API.
AI development
Assistants and agents in the product you have
We add a model where the team already spends time: support and sales chats, answers from the catalog or a knowledge base, and a first pass over requests and documents. An engineer can join your team, or we deliver the project.
Catalog and knowledge
Before it answers, the model retrieves from the catalog, the docs, or another source you name. If that source has no answer, the assistant says so.
Requests and documents
Incoming requests and documents can be read, sorted, and passed on with the fields your process already uses. A person still confirms the step where a mistake is costly.
Models, cost, and deploy
OpenAI, Anthropic, Google, or an open-source model is called from a Node.js or NestJS backend. Evals check a fixed set of questions, and token spend has a ceiling. We test for prompt injection and deploy on AWS.
What is included
- A chat assistant or an agent for support or sales
- Retrieval over the catalog or a knowledge base
- A first pass over incoming requests and documents
- OpenAI, Anthropic, Google, or an open-source model on Node.js or NestJS
- Evals, a token budget, and prompt-injection checks
- Deployment on AWS, with an engineer in your team or a project we run
Questions
Where do we start if it is unclear where AI will pay off?
Pick one place where people already lose time: support replies, catalog questions, or incoming documents. We look at the volume and at what a wrong answer costs before a model is connected. If the case is weak, we say so.
What does a pilot cost, and how long is it?
A pilot is one workflow, not the whole product. The length is set in the brief and is often a few weeks. The price follows who is on it and for how long. We send a number once that workflow is clear.
Can it run without sending data to a third party?
Yes, if the model runs in your own infrastructure. A hosted API from OpenAI, Anthropic, or Google does receive the prompt and the text we retrieve. We agree that choice before the build and do not change it later without you.
How are hallucinations kept in check?
The assistant answers only from the sources we connect, and it can say the source has nothing on the question. We keep a fixed list of questions and compare each answer with those sources. An answer that fails the check is not released as it is.
How is our data protected?
The assistant sees only the sources we connect. Access and deployment sit on AWS, in the account we agree. We test for prompt injection. Your data is not used to train a model unless you ask for that as a separate task.
Can this be added to the product we already have?
Yes. The model is called from your current backend and shown in the interface you already ship. We do not replace the product with a new one. The same people can stay for the next workflow.