03 / Name
The heart of craft.
For us, Obinka means the heart of craft. It is a reminder to care about the thing we make and the way we make it.
Obinka / Products
Better ways to build useful software.
Obinka develops technology products and researches recursive self improvement: how systems can help improve their own capabilities and the methods used to build them.
Our first six products are in development
01 / Portfolio
What we are building
Tools for making, evaluating and using AI systems, alongside products for everyday work and learning.
-
SlopChopper
In developmentStatic analysis for AI-assisted code
Examines code patterns and records each finding with its evidence and limits, so a reviewer can decide what to do next.
Read about SlopChopper -
Paretomax
In developmentAgent harness research lab
A planned lab for comparing and improving agent harnesses, with controlled experiments, candidate history and the evidence behind each result.
Read about Paretomax -
Naagu
In developmentNotes to private podcast
Turns notes and articles into narrated audio with a private podcast workflow, so your reading can have a listening life.
Read about Naagu -
Jecoha
In developmentCoding harness
A local coding harness for longer-running agent work, with tools for inspecting the decisions behind it and testing changes to them.
Read about Jecoha -
Anneta
In developmentExecutive assistant agent
A personal executive assistant agent designed to bring tools, memory and ongoing tasks together for one owner.
Read about Anneta -
Acoolit
In developmentEvaluation framework
Brings agent observations, cost and reported outcomes into view, so the evidence behind a proposed improvement can be inspected.
Read about Acoolit
02 / Approach
The method is part of the work.
We want to build products that are useful in practice and improve the systems that help us make them. Recursive self improvement is the research direction behind that ambition.
01 / WHAT IMPROVED
Name the change
A prompt, a tool, a harness and a model's weights are different objects of change. Each needs its own description before it can be judged.
02 / HOW WAS IT CHECKED
Show the evidence
A changed prompt or a successful task is not enough on its own. The comparison, the conditions and the limits of the evidence belong with the result.
03 / DOES IT HOLD
Test beyond the task
The harder question is whether a gain holds on work that did not guide the change. We want that question asked before a change is kept.
Illustrative method, not a record of measured Obinka performance. Our research site explains the questions in more detail.
04 / Updates
Follow the work that interests you.
Development updates are being set up. Each product page states where its work stands in the meantime.