Private AI on Your Own Hardware

The Problem

TourneyHunter lists thousands of tournaments collected from organizer websites. Organizers move dates, cancel events, and change divisions without telling anyone. To keep the site trustworthy, every listing needs to be checked against the organizer's own page, over and over, all season long.

That's a reading job: open the page, find the dates, compare them to what we have. Paying a cloud AI service to read thousands of pages every night adds up fast. And for many businesses, sending their documents to a third party isn't an option at all.

The Solution

We run the AI ourselves, on a Mac Studio with 128 GB of memory. It hosts open-source models up to 72 billion parameters, the size class that was cloud-only a few years ago. Every night at 11pm it works through the listings on its own.

It uses two models. A fast 30B model reads every page in about a second and proposes changes. A slower, more careful 72B model re-reads any page where a change was proposed, and nothing is applied unless both agree. Anything ambiguous gets flagged for a person instead of guessed at. A run summary lands in the inbox the next morning.

How We Picked the Model

We didn't pick a model by reputation. We built a test: hundreds of real organizer pages with known right answers and known errors, and ran every candidate model against it, including Anthropic's cloud models for comparison.

The number that mattered most wasn't speed or how many errors a model caught. It was how often it invented a date that wasn't on the page. Several fast models confidently made up dates for pages that only said "end of March." They were disqualified. The 72B model matched the best cloud model on honesty and ran entirely on our own machine.

How We Thought About It

Local AI isn't automatically better. It's better for jobs that are high-volume, repetitive, and private, where you can measure whether the answers are right. For the occasional hard question, a cloud model is still the right call, and the system falls back to one automatically if the local machine goes quiet.

The first night it made 15 corrections, and 13 of them had to be undone. We tightened the rules. Since then it makes zero to two corrections a night and none have needed undoing. That's the part people skip: measure, tighten, measure again.

The Result

• 1,000–1,600 organizer pages checked every night in about 20 minutes
• About 7 cents of electricity per full pass, with no cloud AI bill
• Two-model agreement before any change: zero bad corrections since the first week
• No listing data sent to an outside AI provider
• The same setup works for any business with private documents to read: contracts, invoices, intake forms