Kendall Morris
I build simulated workplaces where AI agents are trained and graded: task environments in Harbor and Terminal-Bench format, scored by deterministic graders. No AI judges. Every task passes or fails against ground truth.
Sample environments
Close Ledger
Accounting close operations. Six tasks: bank reconciliation repair, close checklist completion, journal entry correction, fixed-asset register cleanup, accrual verification, and a policy changeover with an effective-date test.
Matter Desk
Legal matter and docket operations, corporate and document workflows only. Six tasks covering matter intake with conflict checks, closing checklists, document register repair, docket deadlines, a filings gap audit, and a retention policy changeover.
Payables Desk
Finance back office: accounts payable and invoice processing. Six tasks covering three-way invoice matching, duplicate invoice voids, approval routing at set tiers, disbursement reconciliation, cost center repair, and a vendor terms changeover.
Validation
Every sample environment is validated before it is shown to anyone:
- Reference (oracle) solutions score 1.0 across five runs on every task, with identical rewards run to run.
- An empty agent scores 0 on every task. Nothing passes by accident.
- Anti-cheat probes (hard-coded answers, grader tampering) fail as designed.
- Tasks are generated from seeds, so variants are reproducible and free of public contamination. The demo pack and sellable inventory are kept strictly separate.
Working together
I build custom environments and graded task packs to order for AI labs and data vendors, and I take on task authoring work through builder platforms. If your team needs operations tasks with graders you can trust, I would like to hear what you are measuring.
Contact
Email: mrkendallmorris@gmail.com