HYPESCRAPER Post yours

← Reports

EXAMPLE — written by the editors. It shows the format. It never counts in votes, numbers or on the Tower.

Data & analytics · Self-hosted open model

Small self-hosted model replaced a big one for ticket tagging

● Worked

What happened

We tagged incoming tickets by product area with a large hosted model. It worked, but every ticket cost more than it needed to. We tried a small open model on one spare server with 1,000 labelled tickets as a test. It matched the big model on 96% of them, and the misses were tickets humans also argued about.

What it cost

From about €1,800 a month to one server we already had, plus a day of setup and a few hours a month of care.

What we fixed

We kept the big model as a fallback for tickets the small one is unsure about. That is about 5% of them.

The lessonTest a small model on your real data before you pay for a big one.

Reality numbers

Team
51–200 people
Monthly AI spend
€100–1k
Still running after 90 days
yes

Got a question about this? Ask engineers who already did it. 2 answers in 7 days or your money back.

Ask the field — free for now