How it works

AI does the trial and error. A fixed check decides.

Every idea is a ticket. Every try ends as a pull request or a written-up rejection. Nothing reaches your main code unless it is proven better, and nothing merges without your review.

What changes

What you get instead.

AI does the trial and error. Your engineers decide what to try and what to merge.

Before: A senior engineer tied up by every idea

Your engineers get their time back

AI studies your code and each idea, writes the code and measures it. Your team reviews results instead of producing them.

Before: “It seemed faster”

Gains you can trust

Old and new code run on the same problems, several times each. A change counts only if it is clearly better, not lucky, and nothing else got worse. A fixed check decides, not the AI.

Before: A cleanup nobody dares to start

Cleanups that are safe to merge

A cleanup only goes through if every test passes and results get no worse.

Before: Failed ideas vanish; know-how leaves with people

Every failure makes the next try smarter

When an idea fails, its code is thrown away but its notes stay in your repo. The next idea starts from what was learned.

Log of triesexample
012 tune a setting to the number of machines dropped
013 new rule for which job goes next dropped
014 reorder the jobs that set the finish time kept
015 same, and favour jobs that are due soon kept
Why try 013 failed, passed on to try 014: “the rule for picking jobs isn't the problem; look at the jobs that set the finish time.”

The failed try is what pointed the next one the right way.

How a try works

Every try takes the same five steps.

Each step leaves notes in your repo and updates your ticket, so you can always check what happened and why.

  1. 01
    Your team or ours

    Add an idea

    An idea is a normal GitHub ticket. Your team and ours share one list.

  2. 02
    AI

    Read up

    It studies your code and the idea, then writes a plan. It also looks at why the last idea failed.

  3. 03
    AI

    Build and measure

    It writes the code, then builds, tests and measures it on a separate copy of your code.

  4. 04
    Fixed check

    Decide

    Plain code compares old and new on the same problems. No AI gets a vote.

  5. 05
    You

    Review or learn

    A win becomes a pull request for you to review; your team decides what gets merged. A failure leaves notes for the next idea.

Seven kinds of improvement · each with its own check
KindWhat it changesCheck
search Smarter ways to look for better answers clearly better results
settings Tuning the settings you already have clearly better results
rules Better rules for building and choosing answers clearly better results
cleanup Tidier structure, same results all tests pass · nothing worse
new rules Handling a new kind of business rule passes new test problems
splitting Breaking big problems into smaller ones passes new test problems
inputs New inputs and outputs passes new test problems
Over a pilot

Your best result only moves up.

Best result so farillustrative · shape only

Your best result only moves up when a change wins.

result ↑ tries →
Not customer data. It shows the idea: a failed try never makes your result worse.
Built on published research
  • FunSearch: AI-written packing rules beat the classic ones
  • EoH: AI improves its ideas and its code together
  • ReEvo: AI learns from written feedback on past tries

Book a call →