Owner Command Center

The Problem

The owner knew the business was profitable. He just couldn’t prove it — not quickly, and not cleanly.

Revenue came through one bank account. Expenses went out through another. Credit card charges lived in a third. Payroll ran through a fourth. And none of these systems had ever been introduced to each other.

Every week, someone would download CSVs, open Excel, and start the ritual: copy, paste, categorize, format, repeat. By the time a P&L was ready, the data was already stale. And if the owner wanted to know something specific — say, how much was spent on cleaning supplies last quarter — that was a whole new project.

The business was doing $735K a year with 22 cleaners. But the owner was making decisions the same way he did at $200K: spreadsheets, memory, and hope.

What I Built

A financial ingestion engine that processes everything automatically.

The system pulls raw exports from all five accounts — deposits, expenses, payroll, credit cards, and payroll register — and runs them through a three-tier classification engine:

1. Overrides — exact matches for one-time exceptions (that weird charge on March 3rd that doesn’t fit any category)
2. Learned maps — 500+ patterns the system has seen before (“COSTCO WHOLESALE #482” → Cleaning Supplies)
3. Pattern rules — regex fallback for anything new

If the system genuinely can’t figure something out, it asks. Once. Then it remembers forever.

Everything lands in a SQLite database — the single source of truth. No more spreadsheets as databases. No more “which version is the latest?”

Three P&L views from one dataset:

• Current Operations — everything, including owner salary and family expenses
• Buyer View — only what a new owner would inherit (the business was exploring a sale)
• Buyer + Normalized Ops — buyer view with a market-rate salary plugged in

The owner can run finance summary and get a full P&L in seconds. Not hours. Seconds.

The Parts Nobody Sees

The classification engine handles the ugly stuff that makes financial data messy in real life:

• Deduplication via SHA256 hashing — re-downloading the same CSV doesn’t create phantom transactions
• Pending transaction filtering — bank exports include charges that haven’t posted yet, and their descriptions change when they do
• Per-employee role mapping for payroll — cleaners go to COGS, office staff to OpEx, owner draws get excluded from buyer view
• Credit card payment detection — so a Capital One payment doesn’t show up as both an expense and a card transaction

None of this is glamorous. All of it is the difference between numbers you can trust and numbers that look right until someone asks a question.

Beyond the Finances

The platform grew beyond just money:

• 28 SOPs covering every scenario from lead intake to service recovery — searchable from a single command
• 149 real-world scenarios contributed by three department heads, tagged by category, linked to relevant procedures
• Job performance metrics pulled directly from the scheduling system — revenue per cleaner, square footage rates, job type breakdowns
• A natural language query layer — the owner types “what did we spend at Home Depot last quarter?” and gets an answer, not a blank stare from Excel

The Result

The owner went from waiting days for a financial picture to getting one in seconds. The Monday morning data compilation ritual — the one that took someone 2-3 hours every week — is gone.

But the real value wasn’t the time saved. It was the questions the owner could finally ask. Which cleaners are actually profitable? What’s our real gross margin after normalizing for deep cleans vs. regular service? If we sell this business, what does the P&L look like without my salary in it?

Those questions used to require a consultant and two weeks. Now they require a command and two seconds.

Stack: Python · SQLite · pandas · openpyxl · Claude API · FastAPI · Next.js