auxls — Stop guessing which prompt works

01 — the problem

The most important code you never test

Somewhere in your product, a prompt is answering customers, writing code or making decisions on your behalf. It was probably written in an afternoon, tweaked when something broke, and never measured since.

That’s a risk, because prompts are fragile. Change one sentence and resolution rates move. Add an example and the tone shifts. Swap the model underneath and yesterday’s best prompt quietly gets worse. Most teams find out from a complaint, not a dashboard.

You would never ship code this way. *Why ship prompts this way?*

02 — how it works

From a prompt to a system you can improve

(1)Describe
Describe the goal.Tell auxls what success looks like (a resolved ticket, a passing test, a completed sale) and the scenario your prompt works in.
(2)Bring
Bring your prompt.Add the prompt you use today, along with any evaluation results you already have.
(3)Sharpen
Let it sharpen.auxls breaks the prompt into tunable parts, tests variations against your goal, and shows you what works and why.

03 — founders

Why us

Two brothers. Rocco has already used the methods behind auxls on real ML models. Ralf has been turning attention into revenue since he was 11.

co-founder · engineering

Rocco Liiva

Mathematics student at the University of Tartu and software engineer. Sets problems for Estonia’s national informatics olympiad. Builds the auxls engine.

ResearchIndependent high-school study applying attribution and feature removal to money-laundering detection models. Supervised by Reiko Rist, Head of EMEA AML Transaction Monitoring at UBS.
2ndEstonian team at NCPC 2025, the ICPC Nordic contest. Qualified for NWERC.
#2 / 4,909Second-highest score nationally on the University of Tartu academic test.
1,000+Clients on the production software he ships as a software engineer.

LinkedIn ↗

co-founder · go-to-market

Ralf Liiva

Founder since 11. Has raised money, sold out shows and sold out a product launch. Runs distribution and sales at auxls.

2,000+Orders waiting at Käpavägi, his subscription D2C brand, after the first run sold out.
15,000+Tickets sold through Back2Back, including a record sell-out for an independent show in Estonia.
Age 14Raised outside investment to fund artist bookings and production.
1M+Followers across the social media network he built and monetized from age 11.

LinkedIn ↗

04 — engines

Two engines, one goal

engine a

Your prompt, broken into dials

auxls reads your prompt the way an engineer reads code. Tone, structure, instructions, examples, constraints and level of detail each become a parameter you can adjust on its own. Instead of rewriting and hoping, you get controlled improvement: turn one dial, measure the result, keep what wins.

engine b

Evaluation that gets smarter

A score is only useful if it reflects reality. auxls pairs LLM-as-a-judge scoring with what actually happens in production: customer ratings, resolution rates, conversions, drop-offs. It uses the evaluation results you already have to calibrate its judges. Over time, the scores stop being opinions and start predicting outcomes.

05 — statistics

Statistics, not vibes

Every recommendation is backed by advanced statistical analysis built to separate real improvement from noise.

a.
Attribution pinpoints which parts of your prompt drive results, and which are dead weight.
b.
Feature removal strips components out one at a time to prove what each one actually contributes.
c.
Interpretation-guided revision turns those findings into targeted rewrites, not random variations.
d.
Sensitivity checks test how fragile a prompt is to small changes in wording, input or model, so you ship versions that hold up.

Every comparison is tested for significance. When auxls says a version is better, it means better.

06 — shipping

Ship only what’s proven

Every prompt version is tracked, scored and comparable side by side, with a full history of what changed and why it mattered.

When a new version beats the one in production, auxls pushes it live through a direct API connection, automatically or with your sign-off. No copy-paste, no redeploy. If it doesn’t clear the bar, it doesn’t ship. And every previous version is one step away if you ever need it back.

07 — who it’s for

Built for teams whose prompts matter

If a prompt touches your customers or your codebase, it deserves the same rigor as everything else you ship.

01
Product teams
Running customer-facing assistants, support agents and AI features.
02
Engineering teams
With LLMs inside code generation, data pipelines and agents.
03
Business teams
Automating content, classification and decisions at scale.

Your best prompt hasn’t been written yet

auxls finds it, proves it and ships it.

Request early access →