Case study · ZYLITIX · Mar 2026 — present
AI tax-prep automation for CPA firms. I own the AI backend: from a shoebox of client documents to a first-reviewed return.
DTAXO reads a client's source documents, builds the workpapers, preps and first-reviews the return, then RPA keys it into the firm's own Drake or CCH Axcess. The firm's people do the one thing only they should: the final sign-off. Every AI step in that pipeline runs on a backend I built.
A CPA firm's season is a mountain of paper: W-2s, 1099s, 1098s, K-1s, 80-page brokerage statements, sideways scans, handwritten notes. Someone reads all of it, builds workpapers, preps the return, reviews it, and keys it into the tax software by hand.
The cost is skilled hours, and the failure mode is a corrected 1099 in March that forces the whole chain to be redone. Firms cannot hire their way out of a three-month spike.
The AI backend for the entire pipeline. Documents come in through the portal; my system classifies them, extracts structured values against tax-form schemas, and reports documentation completeness so the firm knows what is still missing.
From those values it builds the workbook and leadsheet, then preps and first-reviews the return. Preparer decisions that normally need a human are made by the AI, flagged, and documented with the reasoning.
When a late or corrected document arrives, the update flow re-runs extraction, workpapers and the return without starting over.
Upload → documents received and classified → Workbook → Leadsheet → Return Automation. Each return moves through those states in a live task pipeline the firm can see.
The last step is RPA: the finished, first-reviewed return is keyed field by field into the firm's own Drake or CCH Axcess instance, where a human does the final review and signs off.
Everything AI in the backend is mine: ingestion that tolerates any document condition, extraction into structured tax data, workpaper generation, the AI prep and first-review logic, and the re-run flow for updated documents.
The portal, roles and RPA entry are the product surface around that engine; the engine is what I own.
Python for the backend and pipeline orchestration. LLM-based document understanding and extraction chosen because tax documents arrive in every layout and condition; a fixed template parser would break on the first sideways scan.
Per dtaxo.com: 8,000+ returns processed through the platform, 119 CPA firms registered, 2,800+ client entities managed, and firms live on their first return within 24 hours of signing.
In tax, a confident wrong number is worse than no number. The system had to be built so every value traces back to a source document.
Codifying judgment beats automating keystrokes. Once the AI could explain why it made a preparer decision, reviewers trusted the pipeline.
Your version of this
Tell me what your team does by hand. I'll tell you what a system for it would look like.
Everything I know about putting AI into a business, built from scratch in the open: a team of specialist agents that triages, extracts, reconciles, replies and follows up, asks a human only when policy says so, and proves every number back to its source.