I built a system that turns tax law from 100 regions into executable rules

a day ago

TRAM is Sphere’s AI-native tax research system for global indirect tax, automating tax determinations.
It ingests and interprets tax legislation, builds product taxonomies, and determines taxability, rates, and sourcing rules.
TRAM enables rapid expansion into new regions and products with high accuracy through expert-in-the-loop review.
Traditional tax research is labor-intensive, costly, and error-prone, inhibiting expansion.
TRAM operates like a tax researcher, sifting through tax laws to determine product taxability and rates.
It continuously monitors tax authority updates to ensure compliance with the latest rules.
TRAM’s outputs are reviewed by tax experts before being implemented, improving speed and accuracy.
The system improves over time by learning from expert feedback.
TRAM’s architecture includes ingestion, normalization, indexing, reasoning, and expert review steps.
It uses hybrid dense and sparse indexing for accurate retrieval of tax authority materials.
TRAM’s reasoning models generate structured tax proposals with citations and reasoning summaries.
Human experts review and approve TRAM’s proposals, providing feedback for continuous improvement.
TRAM has achieved over 90% accuracy and reduced median expert review time to under 10 seconds.
Future plans include expanding TRAM to cover input tax, withholding tax, and tariffs.

Hasty Briefsbeta