Let’s Talk
#34Sports and fitness / Workforce scheduling20260 of 98 replayed placements went to someone not cleared

CoachCover: Cleared Cover and a Plan B When a Swim Coach Calls Out

Built in-house

Retailbox R&D: built in-house for the Vultr Agent Rush hackathon, 2026

An agent for swim schools: a coach sends a voice note that they can't come in, and it finds a replacement who is cleared to teach that lesson, asks them, and keeps a record of why each person was or wasn't asked. If nobody can come, the manager gets a plan B.

PythonPostgresVultr Cloud ComputeVultr Serverless InferenceWhisperModel Context ProtocolA2A

What was the problem?

When a coach writes at 08:40 that they can't make the 09:30 class, the manager phones around for someone with the right certificates who is cleared for that program. Role, every required certificate and its dates, program and level, weekly hours, rest between days, overlapping bookings and the children-per-instructor ratio all have to hold.

How did Retailbox solve it?

A coach sends a voice note in Russian, Spanish or English. Whisper on Vultr turns it into text, a model translates it, and the agent finds the lessons it affects. The code decides who may take a lesson, and the models only transcribe, translate, read certificate cards and write messages. Offers go to people the rules allow, three at a time, and the rules run again when someone says yes. For a private lesson the parent is asked first. When nobody can come, the manager gets plan B: a reserve coach, overtime, moving the lesson or cancelling it. Approval is tied to a hash of the facts, so a plan whose facts changed is refused. A school can borrow a coach from a sister pool over A2A once both managers agree. A manager's AI assistant can ask who is allowed and request cover over MCP, but approving plan B or answering for a family is refused by policy. A nightly check opens cases before the morning, and an inspector's folder is a PDF one click away. Everything runs on one Vultr VM, with four models on Vultr Serverless Inference.

The school, staff and certificates in the demo are made up. In the replay, how families answer is assumed (80% yes, 10% no, 10% no answer), not measured, and nothing was tested at a real school. The hosted demo ran from 7 to 10 November 2026 and has been taken down; the screens below are from its recorded live run.

Facing the same problem? These are the services that cover this kind of work: AI agent development services.

What does it look like?

CoachCover case page for a 09:30 preschool swim class, covered at 08:45: the voice note in Russian with its English translation, the shift identified, 40 staff checked and 5 allowed with the reason for each refusal, and a Proof panel listing 5 model calls and their cost. (opens the full-size image in a new tab)
The case page: every step with its time, the rule behind each refusal, and the model calls with their price. School and staff are fictional.
Part of a CoachCover case page: offers declined or unanswered, a message flagged for the manager, an approval refused because the plan changed after the reserve coach called out, and a new plan B waiting for approval with overtime, a makeup lesson or a credit. (opens the full-size image in a new tab)
When the reserve coach also calls out, the plan the manager was shown is refused because its facts changed, and a new plan B waits for approval. Fictional staff.
CoachCover network page: one pool's agent asks a sister pool's agent over A2A for a swim instructor, the lender's own rules check the candidate again, both managers approve, and the case shows the loan approved. (opens the full-size image in a new tab)
Borrowing a coach from a sister pool over A2A. Each agent acts only within its own pool's policy, and nothing moves without both managers.

What were the results?

  • ✓On 50 replayed mornings of a synthetic school, a group chat with a phone tree placed someone not cleared for the shift in 45 of 98 replacements; CoachCover did so in none
  • ✓11 of 11 private lessons went to another coach without asking the family in the group-chat replay; with CoachCover, none did
  • ✓The morning case on Vultr took 5 model calls and cost about a tenth of a cent (0.107¢ in the run shown below)
  • ✓Certificate cards are read by two models independently: on 30 synthetic card photos the pair accepted no wrong date and sent 8 of the 30 cards to a person

Have a similar problem?

30-minute call. Free. We'll tell you honestly if we can help.