Your algorithm, measurably better.

AI agents try changes to it. A fixed check keeps only the proven wins, as pull requests.

Book a call
Fig. 1 — finish time example data
start ▼6.8% ▼3.1% ▼4.7% ▼4.3% ▼1.4% ▼ 18.8%
Every merged pull request lowers it. Failed tries ○ never raise it.lower is better
Why it stalls

Your team knows how to make it better. There's never time to prove it.

  • Every idea ties up a senior engineer.
  • Gains that looked real quietly disappear.

Nothing breaks. Your product just stops getting better.

How a try works

Every try takes the same five steps.

  1. 01 Your team or ours Add an idea
  2. 02 AI Read up
  3. 03 AI Build and measure
  4. 04 Fixed check Decide
  5. 05 You Review or learn
What reaches your teamexample data
try 022 reorder the jobs that set the finish time kept ▼4.3%
try 023 tune the cooling schedule to the problem size refused notes
try 024 same, and favour jobs that are due soon kept ▼1.4%

A kept try is a pull request. A refused one leaves notes. Your team decides what merges.

The full method
Twelve weeks

A fixed-price pilot on your algorithm, with a result we agree up front.

If we miss the result, we keep working for up to four more weeks at no extra cost. If we still miss it, you don’t pay the last half of the fee.

You keep

  • the code, as merged pull requests
  • the notes on every try, kept in your repo
  • your data and test problems, which never leave you
Fit

You need four things.

  1. 01an objective function
  2. 02test problems that look like real work
  3. 03a build that runs itself
  4. 04a fair comparison of old and new

No number to improve, or no test problems? Then Argmax isn't for you. Where it applies

FAQ

Questions

What if it only gets good at our test problems?

Honest answer: the result is only as good as your test problems. Argmax makes that visible instead of hiding it. Old and new code run several times on the same problems, so one lucky run doesn't count. You can lock the test setup so no try can change it. New features must come with new test problems. If your tests don't look like real customer work, that's the first thing we'll talk about.

Who owns what?

You own the code, which arrives as normal pull requests in your own repo, the notes on every try, kept next to the code, and your data and test problems, which never leave you. We keep the method and the tools: the AI agents and the checks, used during the pilot and removed at the end.

Where does it run?

On your own servers, with your own AI account. Your code goes only to the AI provider you pick. Every try runs on a separate copy, and your main code only changes when you merge a win.

What if nothing improves?

We agree the result up front, from your own data. If we miss it, we keep working for up to four more weeks at no extra cost. If we still miss it, you don't pay the last half of the fee. A single failed try changes nothing in your code.

Can my team still steer?

Yes. Add ideas yourself, change the order, and review every change before it merges. Argmax suggests ideas too, but the list is plain GitHub tickets that you control.

What does it cost?

A fixed-price pilot of twelve weeks, with a result we agree up front in the first two. You pay in three parts: 20% at the start, 30% when we agree the result, and 50% when it is confirmed. Scope and price are set on a first call. Afterwards you can keep going month by month, or stop and keep everything.

Next

Bring a number. Leave with proven improvements.

A first call about your number, and a result we could agree on.

Book a call

Half the fee is only due once the result is confirmed. Nothing merges without your review.