Loading
Case study
A financial-analytics SaaS that reads a small business's accounting data and turns it into standardized statements, ratios, forecasts, valuations, and AI commentary that will not invent a number.
BizAnalyzer is the flagship app of 2diff.io, the financial-analytics platform I founded and build through my company, Second Difference Solutions. It connects to a small business's QuickBooks Online account with read-only access and turns raw accounting data into standardized financial statements, ratios, forecasts, goals, business valuations, and plain-language commentary.
I own it end to end: discovery, architecture, backend, the AI layer, billing, and deployment. It is a YMYL surface, where a wrong number is not a cosmetic bug, so it is engineered for factual accuracy before anything else.
The accuracy engineering is where my compliance background shows up in the code. Before I built software full time I spent years in broker-dealer supervision, and the monthly-to-annual reconciliation that guards the pipeline is a detective control with a defined tolerance: it catches a classifier drift in the test harness, before a wrong number ever reaches a customer's chart. It is the same controls discipline, pointed at my own data instead of a book of business.
Small-business owners live inside QuickBooks but rarely get analysis out of it. Turning bookkeeping into decisions has historically meant hiring a consultant or a fractional CFO, which most owners cannot justify.
The hard part is not the dashboard, it is trusting the numbers. Every business names its accounts differently, so two companies with identical economics produce completely different report JSON. And an AI that confidently invents a dollar figure on a financial statement is worse than no AI at all.
The core is a layered financial pipeline: standardize the statements, derive the ratios, manage the periods. It normalizes inconsistent QuickBooks report JSON into one comparable model and is validated against real multi-year datasets, including a check that the monthly figures reconcile back to the annual statement.
Above that sits a multi-provider AI layer over Claude, OpenAI, and Gemini, with an anti-hallucination contract: the model receives a structured financial context and is instructed to omit what it does not have rather than estimate it. The platform runs serverless on Firebase and Google Cloud with Python Cloud Functions, encrypted OAuth tokens, Stripe billing, and a static-export Next.js front end.
The parts worth pointing at
Where the engineering decisions actually mattered.
There is a free demo loaded with sample financials, so you can judge the analysis and the AI without connecting an account.