Skip to content

Pilot kit

Iteration 6 prepares the first pilot (FSD §21). Its exit is the MVP exit of PRD Phase 1: one team uses Zettra daily for a month. This kit lists what to set up, how to run the usability sessions, what to watch, and when the pilot counts as passed.

Before day one

  • [ ] Server: sized with hardware.md and installed with operations.md.
  • [ ] HTTPS: with the customer's certificate, or Caddy's internal CA trusted on the team's machines.
  • [ ] AI mode: chosen with the team's IT: local (a GPU for the recommended tier), bring-your-own-key with the acknowledgement signed off, or off.
  • [ ] Exit check for the chosen model: run app eval --seed on the pilot server (FSD §11.8). Record precision, recall, abstention and the median latency.
  • [ ] Load test: run deploy/loadtest/run.sh 100000 on a scratch install of the same hardware, then remove the scratch install.
  • [ ] Backups: run a first backup from Admin → Backups, then a restore drill into a scratch stack. Show IT where /backups is and agree who copies it off the server.
  • [ ] Accounts: the team's users are created, with project roles and client scopes; each person has opened their setup link.
  • [ ] Registry: the module tree is built for the pilot project (§7): modules, menus, client-specific menus, aliases.
  • [ ] History: 20–50 recent requests are entered as tickets, with reasons and closing decisions, so Ask has history to answer from.
  • [ ] Feedback: a channel exists, such as a chat group or a weekly call, and one named person on each side.

Usability sessions

Run three 45-minute sessions in the first two weeks, one person each (a PM, a developer, a support lead). Sit beside them, give each task as written, and do not help unless they are stuck for two minutes.

Note for each task: - whether it was done; - the time it took; - where they hesitated; - what they said, in their words.

# Task (PRD story) Done when
1 "A client just called about overtime approval. Log their request." (story 3) A ticket exists with client, requester, menu and reason
2 "Find out why overtime approval skips the supervisor for Client A." (story 1) They open the menu's page, or ask, and name the ticket behind it
3 "Ask Zettra the same question in your own words." (story 2) They read an answer and open a cited ticket, or see "Not enough information" and know what to try next
4 "The request is done. Close it." The close dialog is filled and the decision record is confirmed
5 "Move a menu to the module where it belongs." It is moved by dragging or with the parent picker
6 (Admin) "Check that last night's backup ran." They find the time and size in Admin → Backups

Afterwards, ask: - What did you expect that did not happen? - What would stop you using this every day? - On a scale of 1 to 5, how much do you trust Ask's answers, and why?

What to watch during the month

  • Admin → Ask log: the not-enough-information rate, answers slower than 30 seconds, and the questions nobody could answer. These point to missing history or menu aliases.
  • Admin → System status: disk use, failed jobs, the model server.
  • Tickets:
  • the share linked to a menu (target 90%) and the share with a filled reason (target 85%);
  • weak reasons, such as "client request", to coach.
  • Daily use: someone on the team creates, updates or asks at least once each working day.

Exit

The pilot passes when, over one month: - the team uses Zettra on at least 18 of about 22 working days; - the success metrics in FSD §15.5 are at or near their targets, measured from the database; - the team wants to keep it.

Record the results, and the fixes the sessions led to, in FSD §21's Progress paragraph.