Skip to content

01.02 / Build

AI features and agents.

Multilingual assistants grounded in your own knowledge base, agents that take real actions through tools, and the evaluation and guardrails that make them safe to put in front of customers.

Book a callTell us about your project
Spanish and English routes prerendered as static HTMLSanto Domingo Bay
38
A hand on the laptop at a hotel reception: the guest inbox with a reply the agent has already drafted for an early check-in, waiting for a person to approve and send it.

In detail

What you should know.

Empty hotel reception desk at night, a monitor showing a call waveform and transcript beside a bookings table, a desk phone and a brass luggage cart

For example: 38 Spanish and English routes prerendered as static HTML. Santo Domingo Bay

Book a callTell us about your project

What it is

Polargate builds AI features that take on 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 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

What you take away.

Everything is written down before we start, and everything stays in accounts your company owns.

01Deliverables

  1. 01

    Retrieval over your own content in Postgres with pgvector, scoped per tenant and fail-closed

  2. 02

    Multilingual assistant with server-side prompts, streaming answers and a clean handoff to a human

  3. 03

    Tool calling with typed schemas validated on the server, exposed over MCP when several clients need the same tools

  4. 04

    Evaluation set built from real questions, refusals included, run before every release

  5. 05

    Guardrails: identity checks, rate limiting per hashed IP, daily spend caps and hard limits in the database

  6. 06

    Usage and cost logging per model, readable per conversation

  7. 07

    Human approval step for anything irreversible, with the proposed action shown in full

  8. 08

    Handover: code in your GitHub organisation, data in your Supabase project, prompts and evaluations in version control

02How it runs

  1. 01 · Week 1 to 2

    Discovery Sprint

    Interviews with the people who will use the thing, an audit of the systems you already run, a data model, an architecture decision and a scope written in plain language. It ends with a fixed price and a delivery plan, and the document is yours whether or not you continue.

  2. 02 · Week 2 to 4

    Foundations

    Design direction, database schema with row level security from the first migration, environments, continuous integration and the first screens in a place you can open. Nothing is built on a platform you cannot leave: the repository, the backend and the hosting are created in your accounts.

  3. 03 · Week 4 to 8

    Build to the first shipped iteration

    Weekly deployments, a short written update with each one and a running list of decisions. The goal of this stage is a version in production with real data where someone can finish an actual task, not a demo. Scope changes are priced before they are built.

  4. 04 · Week 8 onwards

    Audit, launch and handover

    Several review passes over correctness, architecture, accessibility, performance and security, with every finding checked by a second reviewer before it counts as real. Then the launch, the handover pack and, if you want it, a move to the CARE retainer for maintenance and evolution.

For example: 5 valid rich result items on the home page in Google's test. Santo Domingo Bay

Book a callTell us about your project

Who it is for

We don't take every project.

We would rather tell you now than halfway through. If you recognise yourself on the left, let's talk.

It fits if

  • You need a website, a portal or an app that does not exist yet, or one that has run out of road
  • You want the code, the data and the hosting in accounts your company owns
  • You prefer a fixed price before work starts to an open hourly bill
  • Someone on your side can decide and give the project a few hours a week
  • You care that Google and AI assistants can read it from day one

It is not for you if

  • You are after a cheap template or a site put together over a weekend
  • You want it to live on a platform you cannot leave later
  • Nobody on your side can review screens or make decisions
  • You need people to bill hours without a clear goal

For example: 2,500 guest requests handled in four months. Santo Domingo Bay

Book a callTell us about your project

Timing and budget

How we quote.

No price list and no hourly billing during a build. After a short call you get a written quote for your project.

Timeline
Two weeks for the Discovery Sprint, then four to eight weeks to a first version in production. Evaluation and guardrails are part of that work, not a later phase.
Team
Andrés Ciordia, Chief AI Officer, leads the AI side: which model, which tools and how the agent is tested. Our team builds the feature into your product, with voice, native app or data people when a piece needs them, and Pedro Ciordia, CTO, sets the architecture and reviews the code before it ships.
How the price is set
Fixed price for a first AI feature in production, closed after a Discovery Sprint. Model usage is billed by your provider. Care keeps it running once it is live. Ask for a private quote.
The stack
AnthropicClaudeOpenAIMCPSupabasePostgreSQLDenoTypeScriptReactViteVercelZodGitHub ActionsVitest

For example: 3 min median first response from the concierge. Santo Domingo Bay

Book a callTell us about your projectAsk for your quote

Work

Where we did it.

Real projects. Open each one to see what we built and the figures we measured.

For example: 16 min median time to resolve a request. Santo Domingo Bay

Book a callTell us about your projectAll work

Same line

Other services in Build.

Each service has its own page with what it includes, the usual timeline and the projects where we did it.

For example: 93% requests opened by the guest, without reception. Santo Domingo Bay

Book a callTell us about your project

FAQ

What people ask us.

How much does it cost to add an AI assistant to our website or app?
A first AI feature in production has a fixed price, closed after a Discovery Sprint that settles scope and architecture. What moves it is how many systems the assistant reads from and how many actions it can take. On top of that you pay your model provider for usage, which we log per conversation so the number is never a surprise. Monitoring and evolution afterwards run on a Care retainer. Tell us what the assistant should do and you get a private quote.
How long does it take to get an AI feature into production?
Four to eight weeks after the Discovery Sprint, which itself takes two weeks. That covers a scoped feature: an assistant over your documentation, or an agent that opens tickets and takes orders. Voice, a PMS or ERP integration, or a native app add time, and we say so during the sprint rather than after you sign.
How do you stop an AI assistant from making things up?
By grounding it and by testing it. The assistant answers only from content we have indexed, retrieval fails closed so no context means no answer, and the prompt lives on the server where a user cannot rewrite it. Before every release we run an evaluation set of real questions, including ones that must be refused. When the answer is not in your content, it says so.
Can the assistant work in Spanish, English and Dutch?
Yes, and usually in more. Polargate works natively in English, Spanish and Dutch, and the Estevano concierge answers hotel guests in six languages with content translated per tenant. Language is decided per user, not per site. We run the evaluation set in every language you launch in, because an assistant that behaves in English can still drift in Dutch.
Do you train a model on our company data?
No. Your data stays in your own database and is used at query time through retrieval, not for training. We host in the EU by default, on Postgres with row level security so a query can only reach the data its user is entitled to, and the model receives only the passages needed to answer the question in front of it.
What is the difference between a chatbot and an AI agent?
A chatbot answers, an agent acts. The agent gets a set of tools with typed inputs, each validated on the server, so it can create a booking, open a ticket or update a record instead of explaining how. Anything irreversible waits for a person: our own maintenance agent proposes a code change and a human approves every diff before it ships.

For example: 38 Spanish and English routes prerendered as static HTML. Santo Domingo Bay

Book a callTell us about your project

Get started

Tell us what you want to build.

Websites, apps, custom systems and AI for companies that need them in production in weeks. Fill in the form or book a 30 minute call: a senior engineer answers within 48 business hours with a first take on scope, timeline and price.

Book a 30 minute call
What do you need?
Budget
When

At least ten characters.

Answer in writing within 48 business hours. No newsletter, no automatic sequence.

Rather answer the full brief?

First step

Discovery Sprint

Two weeks to close scope, architecture and price before anything is built.

No price list. You get a private quote, in writing, after a short call.

Then, to build, always at a fixed price

Ask for your quoteHow we quote

Not ready yet?

AI-native digital studio for mid-market companies

POLARGATE, S.L. · CIF B24895005
Calle Electrónica 6, 28027 Madrid, Spain