Can you run Gemma 4 26B-A4B on RTX 4090?

Weekly sync, support tools team
Attendees: Maya (lead), Jonas, Priya, Tom, Elena
Agenda: local drafting assistant pilot, knowledge base cleanup, customer feedback survey, on-call rota, next steps

1. Local drafting assistant pilot
- Maya recapped the goal: agents draft replies with an assistant that runs on the team's own Macs, so customer emails never leave the building and nothing depends on an outside service being up.
- Jonas ran the pilot for two weeks on four machines. Agents used it for first drafts on roughly a third of tickets. Most drafts needed light edits. A few needed a full rewrite, mostly billing questions where the model guessed at policy instead of saying it did not know.
- Priya said the biggest complaint was the wait before the first word on long tickets. Short tickets felt instant. Tickets with a long email thread pasted in took several seconds before anything appeared, and agents read that as the tool being stuck. Two agents restarted the app because of it.
- Tom suggested trimming the thread to the last three messages before sending it to the model. Elena worried that would drop context the customer refers back to, like an order number from the first email. Agreed to test both on the same set of tickets before choosing.
- Jonas said the speed once the reply starts was fine on every machine. Agents read slower than the model writes, so nobody asked for faster output. The only request was a visible sign that the model is reading, so people stop restarting it.
- Memory: the oldest machine, the one with the least memory, could not keep the model loaded next to the browser and the ticketing app. It swapped and slowed down badly by mid-morning. The other three machines had no issue all day.
- Priya asked whether a larger model would fix the billing guesses. Jonas thought not: the model guesses because the policy text it needs is missing or contradictory, and a bigger model would guess more fluently, not less.
- Decision: keep the current model size for now. Revisit only after the knowledge base cleanup, and only if the edit rate on billing drafts is still high.

2. Knowledge base cleanup
- Elena found about forty articles that contradict each other on refund windows and plan changes. These are the likely source of the billing guesses.
- Priya proposed one owner per section, with a review date on every article so stale pages show up on their own instead of during an angry ticket.
- Agreed: Elena owns billing and refunds, Priya owns accounts and login, Tom owns shipping and returns. Each owner archives duplicates, merges the rest, and flags anything that needs a policy decision from Maya.
- Jonas will add the cleaned articles to the assistant's reference folder once each section is signed off, not before. Mixing old and new versions would make the guessing worse.
- Tom asked for a template for new articles. Elena will share the one billing already uses.

3. Customer feedback survey
- Priya shared the quarterly survey results. Response time scores went up since the last quarter, and the free-text comments mention faster first replies more than once.
- Several comments still complain about being asked for the same order details twice. Elena thinks this happens when a ticket moves between agents and the notes field is left empty.
- Agreed: the handover checklist gets one new line, fill in the notes field before reassigning a ticket. Tom will update the checklist and post it in the team channel.
- Maya wants the next survey to ask one question about reply quality, not only speed, so the team can see whether the assistant drafts change how customers rate answers.

4. On-call rota
- Two people asked to swap weeks in November. Maya approved both swaps.
- Tom raised that weekend coverage gets thin when two people are on leave at the same time. Agreed that leave requests go in the team channel at least two weeks ahead, and Maya checks the rota before approving.
- No change to escalation rules. The escalation doc is still correct.

5. Next steps
- Jonas: run the trimmed thread versus full thread test on fifty past tickets and report the wait before the first word and the edit rate for each.
- Jonas: add a reading indicator to the assistant panel so agents can see the model is working on a long ticket.
- Elena, Priya, Tom: finish their knowledge base sections and send Maya the open policy questions.
- Maya: ask IT whether the oldest machine can get more memory or should be replaced at the next refresh.
- Everyone: log any draft that invents a policy, with the ticket link, in the shared sheet.
- Next sync: same time next week. Maya sends the agenda the day before.