The problem
Tax prep isn’t hard because the maths is hard. It’s hard because the inputs are scattered across a year of inboxes, banking apps, photos of crumpled receipts, and PDFs from three different brokers. By the time it’s due, half a Saturday goes to assembling the pile before any actual returns get filled.
What we’re building
Tax Prep is a single drop-zone for every document that might matter at tax time. Forward an email, drag a PDF, snap a receipt — everything lands in a vault organised by tax year and category. Documents are OCR’d, classified (income / deductible / capital gains / super / charitable / medical), and tagged with the relevant amounts and dates. At year-end, one export gives your accountant a complete, structured handover.
The AI angle
Claude does the classification. Each new document goes through OCR, then a structured prompt that returns category, dollar amount, date, counter-party, and a confidence score. Low-confidence items go to a small review queue you triage in batches. Confidence calibrates over time as you accept or correct the model’s guesses — the system learns your specific deduction categories the same way a long-running accountant learns them.
How it’ll be used
- Sole traders and contractors who track receipts in three places and lose two of them.
- Investors with brokerage statements, dividends, and capital-gains events that need cross-referencing.
- Anyone with a complex personal return who wants tax time to be a half-hour handover, not a weekend.
Where we are
OCR + classification is working. Email forwarding, manual upload, and bank-statement parsing are in. The next milestone is the year-end export — a single bundle in the format your accountant prefers (CSV, PDF, or a hand-shake with their software). Expected public beta around the next financial year crossover.