Products
Environments, evaluations, and expert data.
Three ways we help frontier models learn and measure expert work. Each is built with specialists from the field and held to the same standard: an expert would sign off on every task.
RL environments
Real software. Real work. Honest rewards.
An environment is three things built together: a working application filled with synthetic company data, tasks written by experts, and a grader that scores what the agent actually changed.
The world
A working application with an API and a screen interface over the same state, reset to the same starting point for every attempt.
The tasks
Written and owned by specialists, reviewed by multiple experts, and placed at Easy, Medium, or Hard by measured frontier pass rates.
The grading
Exact checks wherever there is a right answer, expert rubrics where judgment decides, and side-effect checks that fail any out-of-scope change.
Attacked before release
Doing nothing, guessing, or gaming the grader must score zero before a task ships.
Environments we run today
Business operations
Partnerships and Business Development
Partner, affiliate, sponsor, reseller, and referral programs, from the deal pipeline to the monthly payout.
Business operations
CRM and Growth
A consumer customer base, lifecycle messaging tools, and simulated customers who respond to what they receive.
Business operations
People and Org Design
Team structures, compensation, bonus plans, KPIs, OKRs, and role training, within budget and pay bands.
Business operations
Finance Ops
Payables, receivables, approvals, the ledger, and the month-end close, where balances have to tie out exactly.
Strategy and reasoning
Imperfect-information games
Poker environments, built with professional poker coaches, that test decisions under hidden information and variance.
Any field
New domains
Built with specialists recruited for each program and delivered in the format your training stack uses.
Evaluations
Measure what matters to you.
The same environments and tasks, run as private benchmarks. Built around the capabilities your team is targeting and kept out of every training set.
Held-out tasks
Tasks that are never released, so results reflect skill and not memorisation.
Human expert baselines
How real specialists score on the same tasks, so model results have a meaningful reference.
Difficulty you can read
Results broken down from routine work to the hardest cases in the field.
Failure analysis
Where models go wrong, how often, and what the expert would have done instead.
Expert data
Data from people who do the work.
Written by practising professionals in the domain, reviewed by multiple experts, and checked against the model behaviour it is meant to change.
Demonstrations
Complete, expert-quality examples of how a task should be done, for supervised fine-tuning.
Reasoning traces
How an expert thinks through a hard problem, step by step, including the checks they run along the way.
Preference data
Expert comparisons of model answers, with the reasons behind each choice.
Critique and correction
Specialists mark exactly where a model's answer goes wrong and write the corrected version.
Domains
Any field your models need.
We recruit the right specialists for each program and build to the format your training stack already uses.
Start with a sample.
Every delivery includes its grading report, frontier model baselines, and documentation.