What we build on, why each one, and which model runs which job.

Claude

The model behind most of the agents and drafting work we build.

It follows long, detailed instructions without drifting, and it is straightforward to constrain with tools and explicit limits. That matters more than benchmark scores when the output goes to a customer.

CCodex

A second engineering agent we use alongside Claude when building.

Two independent builders catch each other's mistakes. We use whichever suits the task and never depend on only one vendor.

LLlama (open weights)

Open-weight models doing the high-volume classification and tagging work, served through DigitalOcean's inference endpoint.

At volume the question stops being which model is cleverest and becomes what each call costs. Sorting thousands of news items does not need the expensive model, and the endpoint speaks the same API as the others, so moving a job between providers is a change of address rather than a rewrite.

Next.js

The framework behind every web product we ship.

One codebase covers the page, the API and the rendering strategy. We have shipped it enough times that the first week of a build is never spent on setup.

Vercel

Hosting, previews and scheduled jobs for our web applications.

Every change gets its own preview link before it goes live, and deploys roll back in one click. For most businesses that is the difference between shipping weekly and shipping quarterly.

Supabase

Postgres, authentication and file storage for the applications we build.

It is standard Postgres underneath, so nothing we build for you is trapped in a proprietary format. Row level security lets us separate data at the database rather than trusting application code to do it.

BigQuery

The warehouse where analytics data lands so it can be queried for years.

Analytics tools show you their window. Exporting the raw data means the numbers are yours, and questions the dashboard cannot answer become one query.

DigitalOcean

Servers for long-running work, and serverless model inference for high-volume AI tasks.

A plain server you can inspect beats a black box when a scheduled job misbehaves at 3am. Their serverless inference also lets us run cheaper models for bulk work and keep the premium model for the decisions that matter.

Docker

Containers for the services that run on our own servers rather than on Vercel.

The same image runs on the laptop and on the server, so a change that worked in testing behaves the same way in production. It is also a boundary we can enforce: in one of our own systems the container serving the public API is built without the trading libraries at all and starts with execution switched off, so that service cannot place an order even if something inside it tried.

Cloudflare

DNS, security and edge workers in front of everything we run.

It removes a whole class of problems before they reach the application: bots, denial of service, misrouted domains. The free tier covers most businesses honestly.

GitHub

Where every line of code we write lives, including yours.

At handover the repository transfers to your account. A build you cannot take away from your vendor is a rental, not an asset.

Resend

Transactional email: confirmations, notifications, reports and follow-ups.

Email that reaches the inbox depends on boring configuration done right. Resend makes the boring part checkable, and its webhooks tell us when a message bounced instead of leaving us guessing.

Telegram

Group and direct-message bots, where an assistant answers in the place the conversation is already happening.

A bot in a group is a different problem from a bot on a website: every reply is public, so what it may say in the room and what it may say in a private message are two separate rule sets, and it has to prove who it is talking to rather than trust a display name. It also costs nothing to stand up and needs nobody's approval, which matters when a messaging platform can withdraw an account without notice.

TTwilio

The carrier layer under messaging that needs a real phone number, rather than an app account: WhatsApp Business and SMS.

The API is the easy half. The hard half is the account: business verification, a registered sender, and a platform that can withdraw the channel with no notice and no appeal. That happened to a WhatsApp concierge we had running, and the useful part was the recovery: the concierge logic sat behind the channel rather than inside it, so a replacement went live within days instead of being rebuilt from nothing. We wire messaging that way now as a matter of course.

Stripe

Payments and subscriptions in the products we build.

It is the payments provider your accountant and your bank already understand. We wire it so the money and the records land in your account, not ours.

Google Analytics 4

Measurement of what visitors actually do, wired with real events rather than default pageviews.

Free, standard, and good enough for almost every business when configured properly. Most setups we inherit measure nothing useful. Ours are built backwards from the five questions you need answered.