AI-first software studio

We build software with agents in the loop — and engineers on the hook.

Design, engineering, infrastructure and operations under one roof. A small team, heavy automation, and no handoffs between the people who design a thing and the people who ship it.

Start a project See recent work

What we build

Four things, done properly.

We would rather do a narrow set of work well than list twenty services we rarely touch.

01 — Product engineering

Web products and real-time systems

Customer-facing applications, internal tools, APIs and the boring plumbing that keeps them up. Mostly Elixir and Phoenix LiveView when a product needs to feel instant, Python where the work is data-shaped, Postgres underneath.

02 — AI systems

Agents wired into real processes

Not chatbot demos. Agent workflows connected to the systems a business actually runs on — with the verification, logging and human checkpoints that make them safe to leave switched on.

03 — Design and brand

Identity, interface and print

Brand systems, UI, and production-ready print artwork made in the same room as the code. No agency ping-pong, and no design that turns out to be unbuildable.

04 — Self-hosted operations

Your tooling, on your own servers

CRM, project management, HR and recruitment, design tooling, monitoring — set up on infrastructure you control, so your operating costs stop scaling with your headcount.

Selected work

Products we own, and platforms we have worked on.

We are not describing an approach we have never used. Everything below is live, in production, and serving real users today.

Blok3 curriculum - three levels of the digital-economy programme

Our own product

Blok3 — learning platform for the digital economy

A learning-management platform with payments behind it: a three-level curriculum on blockchain, MiCA regulation and crypto assets, with per-module learning outcomes and quizzes, enrolment and checkout — published in eleven languages.

LMSPaymentsPhoenix11 locales

Read the case study →

OODA AI home page - Universal AI for Compliance

Worked on

OODA AI — universal AI platform

More than 150 models behind one API — chat, document processing, voice, video, images, agents and avatars — inside a GDPR-compliant boundary. Work we did on someone else's platform, and where a lot of our multi-model experience comes from.

150+ modelsUnified APINext.js

Our own business

Tuatara Sailing

Our charter business, taken end to end: identity, a seven-language booking site, press-ready print with measurable QR codes, and a self-hosted back office.

Phoenix7 localesPrintHRMS

Read the case study →

How we work

AI-first is an operating model, not a feature.

Everyone has access to the same models. The difference is how the work is organised around them.

Scope in days, not quarters

We start by building the smallest real version — deployed, behind a URL, with your data in it. Scoping documents age badly; a working thing you can click does not.

Agents draft, engineers are accountable

Agents do the volume: implementations, migrations, translations, print artwork, test passes. The judgement stays with people, and so does the responsibility. Nothing ships because a model was confident about it.

Verification is the actual job

When drafting gets cheap, checking becomes the bottleneck — so we automate that too. Rendered previews, scanned print proofs, machine-checked translations, real requests against real endpoints. We would rather find our own mistakes than have you find them.

You own the result

Your repository, your servers, your domain, your data. No proprietary runtime you can only rent from us, and no seat licences quietly compounding. If you want to take it in-house, everything is already yours.

Evidence

What that gets you in practice.

A sample of work shipped with this model, rather than a wall of logos.

An education platform live in eleven languages

A three-level curriculum with per-module learning outcomes and quizzes, each locale served under its own URL prefix so every market indexes separately.

Phoenix · 11 locales

A booking site live in seven languages

Phoenix LiveView marketing site and enquiry flow, with a multilingual URL structure, per-locale catalogues and language negotiation for visitors arriving from print.

Elixir · LiveView

Brand and print produced in-house

Identity, flags and double-sided flyers taken from sketch to press-ready files — including scan-verified QR codes and resolution checked against the printer's actual dimensions.

Penpot · print

Print campaigns you can measure

Short tracked links behind every printed QR code, logging scans server-side so a flyer, a flag and a banner can be told apart — without waiting for an analytics vendor.

attribution

A company's whole toolchain, self-hosted

Design, project management, CRM, recruitment and monitoring running on one small server behind a single reverse proxy, with per-service memory limits.

Docker · Caddy

Read the full case study →

Stack

What we reach for.

Chosen because we run them ourselves, in production, not because they are fashionable.

ElixirPhoenix LiveView PythonPostgreSQL Docker ComposeCaddy GrafanaLoki PenpotClaude agents

We run what we recommend. Every tool above is in daily use on our own infrastructure — including the recruitment system behind our job listings. If we have not operated something ourselves, we will say so rather than bill you to find out.

Careers

We hire the way we work.

Open roles are posted on our own self-hosted recruitment system — no CV required for most of them.

View open roles →

Contact

Tell us what is not working.

Send a couple of paragraphs about the problem — not a specification. If it is a fit, you will get a straight answer and a first version quickly. If it is not, you will get that too.

sanin@sanin.ai