Case 03 / 03
Credit model and PDF pipeline
For a SaaS product that helps teachers make teaching documents accessible.
Context
A SaaS product that helps teachers make teaching documents accessible: AI adds headings and alt text, then returns the improved file. It serves individual teachers and schools.
It handles Word, PowerPoint, Excel, PDF and Google Docs links; billing runs through Stripe; folders sync with an LMS.
Decisions
01
Change the display terms: School to Organization, Teacher to User
- Why
- The product sells to schools and enterprises; changing the display layer avoids a data migration.
- Trade-off
- Copy has to stay consistent US English; internal DB and API names still differ from the UI.
02
10,000 tokens = 1 credit; charge only on success, refund a failed retry
- Why
- Teachers need to know the cost up front; a credit is only spent when there is a result.
- Trade-off
- The estimate is final even if actual token use differs; manual adjustments have no transaction log.
03
Drop Adobe and build the PDF auto-tag engine in-house; Adobe quality is the floor
- Why
- Adobe bills per page and has a free-tier cap; building in-house removes that cost.
- Trade-off
- Engineering risk: the engine must not do worse than Adobe across 20–30 real files, measured with PAC 2026.
Outcome
Scope, credit model and PDF pipeline were fixed in the spec; the engine has a list of structures to tag and a measurable quality floor.