Engineering
Why we don't use vector search for retrieval (most of the time)
Embeddings are great for fuzzy. Personal records reward precise. Here's the architecture we landed on after eight months of trying both.
Home / Blog
BlogField notes from a team building memory.
Engineering notes, design decisions, and the occasional opinion. We post when we have something to say — usually 1–2 times a month.
Privacy
"We don't sell your data" is not a privacy policy
A walkthrough of the architectural decisions we made to ensure we couldn't sell user data even if we wanted to.
Design
The hardest screen to design is the deletion screen
Most personal-data apps don't ship a great delete flow because nobody at the company benefits when the user clicks it. We had to flip that.
Engineering
Parsing 18 banks' PDF statements with one prompt
A short story about giving up on regex, then giving up on JSON schemas, then arriving somewhere unexpectedly clean.
Changelog
Changelog · April 2026
FasTrak per-trip attribution, statement parser for HELOC, drawer sightings, and a real fix for the Tesla Fleet API browser OOM.
Engineering
A taxonomy of vehicle documents (and why we needed one)
Service receipts, recall notices, registration, smog, warranty letters, DMV correspondence — they all rhyme but none are the same.
Page 1 of 4
Older →