RL environments

Train agents in the apps each country runs on.

Replicas of the shops, banks and public services people use in each country, built as sandboxes your agent can train in, with graders checked so models can't cheat.

Why country by country

Every country runs on different apps.

Different shops, banks, tax forms and address formats, in a different language. An agent that only practiced on American apps hasn't seen them.

  • Runs today
  • Not built yet
  • Online shops Runs today
  • Banking
  • Government forms
  • Delivery and travel
  • Classifieds
  • Utilities and telecoms
  • Construction drawings Runs today

In every environment

A working replica, and the evidence for every task.

  • A replica, not a screenshot

    A working copy of the app that resets for every attempt, so each rollout starts clean and is graded from the app's own records.

    northwind · demo shop

  • Tasks in the local language

    Prompts in the country's language. Polish first; more languages on request.

    PL

  • Graders that hold up

    Each grader must accept the reference answer and reject doing nothing, then face scripted exploits. One that works rejects the task.

    12 tried · 0 credited

  • Two blind expert solves

    Two experts solve each task without the answer or each other's work. Reviews stay sealed until both are in.

    Queued

  • Ready for your stack

    Exported in HUD's open task format or as Harbor tasks, each with its environment, grader and a provenance manifest.

  • Construction takeoff environments

    US concrete drawings in, quantities out. Designed and checked by a concrete estimator.

    JSONL + COCO exports

How it works

From your country list to a task set.

  1. 01

    Name the country and the apps

    On a call: where your agents will work, and the apps they'll need there.

  2. 02

    We build the replica

    A sandbox that behaves like the app, with tasks drafted in the local language.

  3. 03

    Every task is attacked and solved

    Agent rollouts, grader checks and scripted exploits, then two blind expert solves.

  4. 04

    You get the set and its evidence

    Each task with its report. Pilots of 25, from $200/task.

Said plainly

What runs today, and what we build next.

Today

One demo shop, tasks in Polish, the whole pipeline.

  • Northwind Supply Co., our demo web shop, with tasks written in Polish.
  • A first batch of 23 tasks, each scored by 6 agent rollouts.
  • Grader checks and scripted exploits on every task; 3 of 3 planted broken tasks caught.
  • Construction takeoff environments from US concrete drawings.

Next

The replicas your agents need.

  • Replicas of the apps you name, in the countries you name.
  • Tasks in more languages, solved by native speakers from our expert waitlist.
  • Records from companies that license them to us, through data partnerships.

Tell us which country, and which apps.