Problem
Tax preparation involves reconciling mountains of source documents — W-2s, 1099s, K-1s, brokerage statements, receipts — against the return itself. A missed carryover, a transcribed digit, or an unclaimed deduction is invisible to a hurried reviewer but expensive to the client. CPAs have time to spot-check, not to line-by-line verify every figure, and existing software checks arithmetic, not completeness.
Approach
- Document intelligence: parse heterogeneous tax documents into structured facts (amounts, dates, entity names) with source locations preserved for audit trails.
- Agent orchestration: an LLM-driven review agent plans its own pass over the return: which documents to pull, which line items to verify, which cross-return heuristics to apply (carryover consistency, prior-year deltas, common missed deductions).
- Findings, not vibes: every flagged issue cites the document, the line, and the mismatched value — a reviewer confirms or dismisses rather than re-doing the work.
- Synthetic evaluation set: trained and measured against mock tax data with injected error classes, so detection accuracy (90%+) is a measured number, not a demo trick.
Outcome
The project placed 3rd at the Innovation Lab Competition. More importantly, it was the prototype of what I now build full-time at Juno: agentic tax review with year-over-year comparison, carryover-loss verification, and optimal-strategy checks. The jump from capstone-era prototype to production CPA tooling is the throughline of my career so far.