Scoring buyer intent without lighting money on fire
A two-pass approach: cheap first-pass on everything, careful escalation only on the close calls.
Scoring every social mention with a large model is a great way to run up a bill and learn nothing. Most mentions are obviously noise or obviously hot — you don't need your most expensive reasoning to tell them apart. The money is in the middle.
Filter before you spend
Before anything hits a model, a cheap, deterministic prefilter throws out the obvious junk — off-topic threads, ancient bumped posts, keyword coincidences. In practice this removes 40–60% of volume before it costs a cent.
Two passes, not one
What survives gets a fast first-pass score from a small model. Anything it's confident about — clearly high or clearly low intent — is done. Only the genuine close calls, the 40–70 band where the answer actually matters, get escalated to a more careful model.
Spend your compute where it changes the decision — nowhere else.
Why this works
- Cheap first pass handles the easy majority.
- Expensive second pass only touches the ambiguous few.
- Prompt caching keeps the shared context nearly free.
- You get a 0–100 score on everything, at a fraction of naive cost.
The result is scoring that's both careful and affordable — which is what makes a generous free tier possible in the first place.