labs
Independent software lab Est. 2026

Thoughtful products.
Precise tools.

From everyday rituals to the way we build software.
Small ideas, carefully engineered.

Ideas into useful things.

Selected projects

01 — 08
01

lauta design

Spatial design

Give your bathroom ideas a little room.
A visual space to plan, arrange, and explore in 3D.

lauta.designVisit site
02

steamtrain

Developer tools

Coding agents, working in concert.
Turn complex software work into repeatable workflows.

steamtrain.appVisit site
03

caeto router

Developer tools

A System One router for AI models.
Quick judgement on which model, and how hard it should think.

In developmentRead more
04

fast-code-embed

Open source

Find code that means the same thing.
Lightweight code embeddings, without a GPU or API keys.

githubView repository
05

trama

Everyday life

A calmer path to dinner.
Meal planning that makes the week feel lighter.

In development
06

repertorio

Personal knowledge

Keep what catches your eye. Find it again.
A personal library for scattered inspiration.

In development
07

lupi

Video AI

Something is taking shape, frame by frame.
More when the time is right.

Stealth
08

camelo

Programming language

A new language. A different starting point.
Quietly in the making.

Stealth
03 — In development
Developer tools

caeto router

A System One router for AI models.

Human thinking runs on two systems: one fast and intuitive, one slow and deliberate. caeto router is the fast one. It reads each task, makes a quick judgement about which model should take it and how hard that model should think, then gets out of the way. Deliberation is left to the models.

When a task turns out harder than it looked, the router notices and steps up to a more capable model or a higher reasoning effort, carrying the work so far with it wherever the provider allows.

  1. 01Read A fast assessment of what the task demands.
  2. 02Route A model and reasoning effort, across providers.
  3. 03Escalate More capability only when the work calls for it.

Estimated performance

Preliminary
≤2 pts
from always using a top frontier model, across four families of coding benchmarks
~⅕
of the cost on everyday coding tasks, a few points from the frontier
~60%
lower monthly spend than a single frontier model, on a typical working mix

Estimates, not measurements. They combine published model results, the router’s measured behaviour on public benchmark tasks, and modelled reasoning effort and escalation. Figures vary by configuration; treat them as indicative.

First test results: 27% lower cost on Sonnet High workloads with Anthropic models. Preliminary.

Coming soon

A whitepaper and official benchmark results.