Last week something bad happened. My email provider’s spam filter quarantined a hotel folio. I love spam filters, 99.9% of the emails labeled Junk/Spam are correct, and there’s probably millions of emails I never see.
But… the cost of a false positive now is really high. My hotel folio was mislabeled so my agents & automations didn’t pick it up and couldn’t run my Folio Checker. My folio checker looks for billing mistakes like “valet” l in NYC. My agent knows I very rarely valet a vehicle in the US, but I do have legit valet charges (for laundry) when traveling overseas.
To rescue these emails, I compare the email contents against ACTUAL credit card swipes (via Plaid and AwardWallet) and established related travel plans. Established flights to a city can rescue hotel emails, restaurant bookings, or train reservations. The spammers won’t have this data, unless they’re from Las Vegas (best CRMs in the world).
There’s no LLMs at runtime, just rules, but the rules are written and maintained by an LLM incorporating my explicit signals (e.g. sending back to Spam/Junk). I’ve also got an ongoing automated eval to see if the LLMs beat my rules based system and what should be updated.
Rescuing legit emails from spam/junk based on my credit card swipes and travel plans.
To rescue incorrectly junked/spammed emails, I built a system that compares the email contents against ACTUAL credit card swipes (via Plaid and AwardWallet) and established related travel plans.
