AI features and agents

What it is

Polargate builds AI features that do real work inside your product. Three shapes cover most requests: an assistant that answers from your own knowledge base and admits when it does not know, an agent that acts through tools (it opens a ticket, takes an order, books a table, updates a record), and a classifier that triages what arrives so people only handle what needs judgement. Everything works in the languages your customers actually write in.

How we do it

We start with retrieval, not with the model. Your content is chunked, embedded and stored in Postgres with pgvector, scoped so a question can only reach the data of the tenant that asked, and retrieval fails closed: no context, no answer. Prompts live server-side in an edge function, never in the browser. Tools are declared one by one, with typed schemas validated on the server, so an agent can create a booking but cannot invent one. When the same tools should serve several clients, we expose them over the Model Context Protocol, the way our internal MCP server exposes more than a hundred tools to Claude.

Then comes the unglamorous part, the one that decides whether the feature survives contact with real users. We build an evaluation set from real questions, including the ones that should be refused, and run it before every release. We add identity checks (the Estevano concierge verifies a guest by room and PIN before it acts), rate limiting per hashed IP, daily spend caps and usage logging per model and price. The hard limits live in the database, not in the prompt: our own AutoFix maintenance agent cannot reply to a client or deploy on its own, by constraint, and a person approves every diff.

What you get

A feature in production on your own stack, not a pilot on someone else's platform. The code sits in your GitHub organisation, the data in your Supabase project, the prompts and evaluations under version control next to the code. You get the retrieval pipeline, the tool layer, the guardrails, an admin view of what was asked and what was answered, and the cost per conversation. You also get an honest map of what the feature should not do yet, which is usually the most useful page of the handover.

What it costs

A Discovery Sprint starts at 4,900 euros and ends with scope, architecture, an evaluation plan and a fixed price. A first AI feature in production starts at 9,900 euros, fixed, paid in stages. After launch your model provider bills usage at cost, and the feature can move to a Care retainer from 850 euros per month covering monitoring, evaluation runs and evolution. No hourly billing, no open-ended budgets.