Training grounds for agents that do real work.
Datum builds reinforcement learning environments for business work: working apps, tasks written by the people who do the job, and verifiers built to resist reward hacking.
Partners
| Partner | Conversions | Revenue | Payout |
|---|---|---|---|
| Halcyon Streams | 612 | $48,210 | $12,053 |
| Lumen Creators | 342 | $21,890 | $13,680 |
| Brightline Media | 127 | $11,905 | $6,000 |
| Copperleaf Tech | 88 | $9,370 | $1,874 |
| Ridgeway Sports | 41 | $4,480 | $1,640 |
| Northbeam Podcast | n/a | n/a | $3,500 |
What we build
Real work, real apps, verifiable rewards.
Every environment starts with the people who do the job. Experts map how the work really gets done, and we build the apps, data and tasks from that, so the edge cases, permissions and business rules are already there.
Working business apps
Isolated, resettable sandboxes with an API and a screen interface over the same state, filled with fully synthetic company data.
Tasks written by experts
Practitioners in each field own every task and rubric, from single lookups to long-horizon work that takes judgment.
Verifiers you can trust
Graded on the final application state against an oracle solution, with sealed graders and adversarial testing before release.
Verification
A high reward means the work was done.
We assume the policy will look for the cheapest path to a high score, and design every verifier so that path does not exist.
End-state grading
Final application state is compared to an oracle replay on the same seed. What the agent says it did does not count.
Collateral damage checks
Any write outside the task's scope fails the episode, even when the target change is correct.
Expert rubrics, checked
Rubrics cover judgment only. Experts write them, a judge model scores them, and the judge is measured against expert grading.
Sealed graders
Oracles, expected states and rubrics run in a separate container with no network. The agent only sees the app and its tools.
Adversarial release
A task ships only if its reference solution passes and do-nothing, refuse-everything, random and read-everything agents all score zero.
Held-out splits
Every environment keeps a split that is never licensed, so results can be checked against tasks no model has seen.
| Level | Shape | Target frontier pass@1 |
|---|---|---|
| Easy | Single records, a few tool calls | 60 to 90% |
| Medium | Cross-record workflows, business rules and permissions | 25 to 60% |
| Hard | Long-horizon work with judgment, conflicting sources and policy traps | 1 to 25% |
Environments
The Business Operations suite.
Four specialised apps, licensed together or one at a time, with more verticals to follow.
Partnerships and Business Development
Partner, affiliate, sponsor, reseller and referral programs: deal pipeline, contract terms, performance, payouts and partner communication.
- Revenue share
- CPA and caps
- Guarantees
- Sponsorships
- Deal registration
- Payout reconciliation
CRM and Growth
A consumer customer base with behaviour and purchase history, plus email, push and in-app tools. Agents segment, design journeys and set offers. Simulated customers respond, so results are measured in conversions and revenue.
- Segmentation
- Lifecycle journeys
- Offers and margin
- Consent and frequency
- Simulated customers
People and Org Design
Org charts, team structures, compensation packages, bonus plans, KPIs and OKRs, and role training, all built within budget and pay bands.
- Org design
- Compensation
- Bonus structures
- KPIs and OKRs
- Role training
Finance Ops
Payables and receivables, approvals, bank feed, ledger and month-end close, where balances have to tie out exactly.
- Three-way match
- Accruals
- Approvals
- Month-end close
- Reporting
Imperfect-information reasoning
Heads-up no-limit hold'em environments testing decisions under hidden information and variance, graded by expected value lost against solver output.
- Solver spots
- Known opponents
- Full matches
Custom environments
Task families built to the capability gaps your team is targeting, delivered in the format your training stack already uses.
For AI labs
- Runnable sample under NDA within three business days
- Baseline report on current frontier models
- Exclusive and non-exclusive licenses
- Any task you flag as broken is replaced free
For data partners
- Environments and task volume for your client programs
- Custom task families written to spec
- Generators and seeds included
- Direct line to the people building it
For experts
- Partnerships, growth, finance, HR and more
- Paid work mapping your workflow and writing tasks and rubrics
- Remote and flexible
- Top-ranked experts get more work and higher rates
Put your agents to work.
Tell us what you are training for and we will send a runnable sample.
Or email hello@datum.engineer