November 2024. I'm staring at my screen, comparing two spreadsheets line by line. Transaction 3,247 of 6,000. My client's Xero PayPal account is heavily out of balance, and I'm manually matching transactions like it's 1995.
I'm an accountant with 12 years of experience. My only programming background? A single C++ module during my Astrophysics degree. But sitting there at transaction 3,247, something clicked: this is algorithmic. Why am I doing this by hand?
What I built (10 months later)
Three tools that now talk to each other:
ReconcileIQ - Bank reconciliation at 5,000 transactions per second. PDF statement parsing for 20 UK banks. The thing that started it all.
CodeIQ - AI-powered transaction coding. Upload your PDF statements, go for coffee, come back to categorised transactions ready for posting. Learns from your historical patterns and from anonymised patterns across all users (federated learning - everyone benefits, nobody's data leaves their account).
LedgerIQ - CFO-level analysis from any general ledger export. 20+ modules covering everything from DuPont analysis to working capital management. One click from raw CSV to full financial insights.
The full workflow: PDF bank statement → parsed transactions → AI-coded entries → posted to QuickBooks/Xero/Sage/Pandle → comprehensive financial analysis. End to end.
What actually works
The federated learning surprised me. When the system learns that "AMZN MKTP" maps to office supplies for one user, it can suggest that mapping to others - but using their own chart of accounts, even if category names differ. 90%+ accuracy on recurring transaction types.
Historical pattern matching from your own books is even more accurate. The system learns your specific coding decisions and applies them to similar transactions.
What's still hard
Merchant extraction from transaction strings remains painful. A transaction might appear as "TESCO STORES 3297 LONDON GBR" or "TSC ST3297 02/03" - same merchant, completely different formats.
Multi-currency complexity is real. Which exchange rate, when it applied, how different platforms expect it recorded. Still improving this.
What I learned
Domain expertise matters more than technical skill. I could learn the technology because I understood the problem deeply. Every database decision, every API choice made sense because I knew exactly what bookkeepers need.
Purposeful learning compounds fast. I didn't learn databases in abstract - I learned them because I needed to store millions of transactions efficiently.
The ask
For those building in accounting/finance - what features actually matter to users? What am I missing?
Landing page: https://bankreconciler.app
Genuinely curious what this community thinks. The solo founder accounting-tool space seems weirdly underserved given how painful the manual work is.