AI agents & assistants
Retrieval-augmented generation, or RAG, lets an AI answer from your own documents instead of what the model happens to know. We build it into agents and internal tools: it finds the passages that answer a question, writes the answer from them, shows where it came from, and says so when your documents don't cover it.
We load PDFs, help centers, catalogs and spreadsheets, split them into passages, and index them again when they change, so answers don't go stale.
Embeddings in Postgres with pgvector, combined with keyword search for the names, codes and prices that meaning alone misses.
The model writes only from the passages it was given and cites them. When they don't answer the question, it says so instead of guessing.
Graded cases check that search finds the right passage and that the answer stays true to it, before launch and on every change.
We read real conversations and write down what the agent handles, what it never does and when it hands off to a person.
Tools for your catalog, calendar or CRM, and retrieval over your documents, so every answer has a source.
Graded test cases from your conversations, with a pass threshold. They run in CI on every change, and again when the model or the prompt changes.
A first group of real customers, a console to review conversations and hand-offs, and a second provider that takes over if the main model is down.
To answer questions
An agent that answers your customers on one channel, hosted and looked after by us.
From $1,500 + $250 a month
Setup paid up front, then billed monthly in advance. A 3-month minimum, then cancel any month.
To act in your systems
One agent for one job, wired into your systems and built in your own accounts.
From $8,000
A deposit, then billed by milestone.
RAG, almost always, when the answers live in documents that change. Fine-tuning changes how a model writes, not what it knows today, and has to be done again whenever the facts change.
PDFs, Word and Google Docs, help centers, catalogs, spreadsheets, and databases with an export or an API. Scanned pages need text recognition first, which we confirm while scoping.
In a database in your own cloud account for an AI Agent, or on our hosting for an Essential Agent, kept apart from every other client's. They're used only to answer for you.
An Essential Agent that answers from your documents starts from $1,500 plus $250 a month. RAG over many sources, or wired into your own tools, is an AI Agent engagement, from $8,000, with a fixed price per milestone. Model usage is billed by the AI provider to your own account.
The home page of this site replays a pull request: spec, plan, changes, tests, deploy. This guide walks through that pull request, the one that added AI triage of new leads, and what a person did at each step while agents wrote the code.
A chatbot follows a menu someone wrote. An AI agent reads what the customer wrote and looks up the answer. Both run on WhatsApp, both have a place, and the right one depends on what your customers ask.
A demo shows the answers someone picked. Evals are graded test cases with a pass threshold, and an agent stays off until it meets it. This is how we write them, with the evals behind this site's own lead triage as the example.
These are the four steps we follow to build an AI agent on WhatsApp, with what Meta's platform asks for at each one. It's written for the person who decides whether to build one, with enough detail for the engineer who will.
With us, an agent that answers your customers on one channel starts from $1,500 plus $250 a month. One wired into your systems starts from $8,000, at a fixed price per milestone. Running either costs model usage and channel fees, billed by those providers to your own accounts. This guide explains what moves both, and what to ask any team before you hire them.
Agents that answer, sell, book and triage over WhatsApp, web chat and your own tools.
The spec, the reviews, the tests and the evals behind every service.
A person reads every brief and replies with questions or a first plan.
On this site
The lead triage on this site has 31 graded cases and has to pass 90% of them before it's turned on.