I build reliable tax software and AI development tools, connecting domain expertise, system design, and team leadership to turn complex rules into software people can verify.
Form changes that reach the right people. Extending automated ingestion, version comparison, and the forms viewer so developers can trace a release through to the work it creates.
AI with a verification path. Helping expert teams use source context, compiler checks, and regression results to evaluate generated code.
Clear ownership across two products. Leading federal tax development and serving as the department's AI ambassador, with an emphasis on repeatable workflows and useful tools.
02 / Selected work
Three problems. Working systems.
Built, co-built, and put to use
01 · Change detection
From a form notifier to a production pipeline
Problem / context
Teams downloaded tax forms and compared revisions by hand. The same change could be reviewed several times before reaching a developer.
Architecture / solution
I built the original notifier, then co-built the production platform with engineering colleagues: versioned documents, revision comparisons, change summaries, and links to the development work.
1 day
From detecting revised Form 2441 instructions to shipping the corresponding calculation update.
August 25–26, 2026 · One prelaunch test-workflow example. Federal production launched August 28.
02 · Developer tooling
Give AI the code context and the checks
Problem / context
AI assistants invented syntax for proprietary tax languages and missed dependencies spread across repositories.
Architecture / solution
I built code intelligence with 37 Model Context Protocol tools across 63 repositories, plus build and regression servers. Shared language rules give the model a reference; the real tools check its work.
160 returns
Zero calculation differences in one Form 8801 conversion regression batch.
March 31, 2026 · Print differences remained. This result measures calculation parity in that batch.
03 · Data quality
One form identity across disconnected systems
Problem / context
Form titles varied across registries and work trackers. Fuzzy matching missed existing support and created duplicate work.
Architecture / solution
I drove adoption of a persistent form key and built reconciliation scripts with a reviewable decision report. Approved decisions were applied and checked against the source systems.
332 → 0
Open items remaining in the reconciliation tracking backlog.
Completed August 30, 2026 · Confirmed gaps and accepted exceptions were recorded separately.
03 / Writing
Notes on the decisions behind the code
In draft
I'm developing these notes for review before publication.
Stable identity comes before automation
Why matching records by their display names makes an automated workflow fragile.
2 min read · Draft
Give AI a way to check its work
Source context helps generate a change. A compiler and regression evidence help decide whether to keep it.
Taylor Riley