Berlin · Agent-first

sail.black

Agentic systems, applied AI, and deep tech — shipped so software can act, with humans in the loop where it counts.

Librebase, Taktwerk, Majico, Bureauzilla, The Agentic Book, and Reeldemo.

How we build

Software that runs without babysitting

01

Automate volume

Specialists decide. Software handles intake, diffs, queues, and follow-ups.

02

Ship to learn

Launch, watch real usage, fix what breaks. Repeat.

03

Own the stack

Identity, billing, UI, and infra stay in our repos. No black boxes.

ResearchMore coming soon

Agentic AI, Simulations, Deep Tech

Plain-language articles on our papers and preprints: chemistry, drug discovery, Li, and Reeldemo signal healing.

2026-10-27sail.black

How do you stay in control of many SI agents at once?

Many SI agents acting for you raise four questions: who allowed this, what may it do, can bad input spread, and can you stop it. Our draft plan to answer them.

Read
2026-10-27Reeldemo

How to remove the click in a wavetable loop

A plain-language guide to DenoiseOpt, a search method that repairs the click where a wavetable loop wraps around. With the scores, limits and paper.

Read
2026-07-27Reeldemo

Signal discontinuities matter beyond music

The seam repair built for audio loops was also tested on bearing vibration, heartbeat and synthetic machine cycles. Here are the scores and the limits.

Read
2026-07-27Li

Li: a programming language for fast, checkable software

Li is a sail.black systems research language for code that must run fast and stay checkable, such as simulations and core calculations. What it is and is not.

Read
2026-10-27

How do bacteria close the rings of antibiotics? A new idea

A new preprint proposes that protons hop along chains of water when polyketides such as tetracyclines close their rings. What the idea says and how to test it.

Read
2026-10-27

Can fluorine form triple bonds? A plain-language look

Fluorine is taught as a pure electron taker. A ChemRxiv preprint asks if it can form a triple bond with metals such as manganese and nickel, finds triple-bond character in Ni(III)F, and shows where standard methods break.

Read
2026-10-27

CARATE: teaching computers to read molecules as graphs

CARATE turns molecules and proteins into graphs and learns a compact representation for property prediction. What it does, the benchmarks, and the limits.

Read
2026-10-27

CACHE #1: can computers find drugs for a Parkinson's target?

In CACHE #1, 23 teams used computer methods to pick molecules for the Parkinson's protein LRRK2. 1,955 were tested in the lab. What worked and what did not.

Read
2026-10-27

Watching drugs move: Markov models of protein interactions

Drug design often uses a frozen snapshot of a protein. This FU Berlin thesis models how drug-protein contacts change over time, using simulations and Markov chains.

Read