Aegis BI: Building an AI Financial Command Center for Small Business

Problem

Small and mid-sized businesses have real financial data but no live picture of it. The numbers sit in spreadsheets and accounting exports; runway, revenue trajectory, and risk are reconstructed manually, quarterly at best. Enterprise BI platforms assume a data warehouse, an analyst team, and a six-month implementation, none of which these companies have.

Challenge

The hard constraint was onboarding: if setup requires a data pipeline project, the target customer never gets to value. The platform had to accept the data businesses actually have, a spreadsheet, and produce an executive-grade command center from it, automatically.

A first upload solves the first day and nothing after it. A dashboard that only knows what you last handed it goes stale on a schedule nobody agreed to, so the platform also had to read the workbook where the team already maintains it, in OneDrive or Google Sheets.

The second challenge was trust. Forecasts and AI recommendations are only useful if an operator believes them, which means visible reasoning, honest uncertainty, and a human always in the loop.

Solution

Aegis reads the files a business already keeps. Upload a workbook, or connect the one living in OneDrive or Google Sheets, and the platform maps columns, builds a KPI command center (revenue, expenses, net, coverage, margin, runway), and layers on forecasting, anomaly signals, and what-if scenario modeling. Connected sources are read where they sit, so the dashboard tracks the file rather than a copy of it.

It also leaves the browser. Aegis shipped first as an installable progressive web app and is now on the App Store as a native app for iPhone, iPad and Mac. That was a deliberate decision rather than a responsive-design afterthought: the moments these numbers settle an argument are meetings, site visits, and airports, none of which happen in front of the office desktop.

On top of that sits Glaukos, the built-in AI analyst. Glaukos reads the live dashboard context, runs risk-first analysis, and answers questions in plain English: briefings and recommendations with the reasoning shown, never silent actions.

Technical Implementation

The platform is Python end to end: a Dash (Plotly) front end with analytics running in-process against a Pandas services layer, and a FastAPI backend hosting the AI agents. The defining decision is where data lives: in the user's browser. Uploads are parsed in-session and stored in the browser's on-device database (IndexedDB); the backend is a stateless calculator that persists nothing.

Glaukos and the column-mapping agents call OpenAI's models with careful context engineering: structured dashboard state (metrics, trends, flags) travels with each request, so the analysis references the customer's actual numbers instead of generic financial advice. AI features send data out only with the user's explicit consent, with a strict local mode that forbids it entirely.

• Dash (Plotly) command-center front end, analytics in-process via Pandas
• Stateless FastAPI backend: datasets travel with the request, nothing persisted
• On-device storage: the browser's IndexedDB holds each user's source library
• OpenAI-powered agents (mapping, Q&A, briefings) gated behind explicit AI consent

Results

Aegis BI runs live in production on the web and as a native app on the App Store for iPhone, iPad and Mac, with the dashboard open to explore and companies onboarded directly. It serves as the flagship proof of our decision-intelligence approach: the files you already keep in, command center out, with an AI analyst that operators can interrogate rather than blindly trust.

It also compounds the consulting side of the business: the platform is the productized form of the dashboards, forecasting, and AI-solutions work we sell, which means the service offering is demonstrated rather than described.

Lessons Learned

Onboarding friction dominates everything: meeting businesses where their data already lives, first by reading their spreadsheets and then by connecting to them in OneDrive and Google Sheets, mattered more than any individual feature.

Where an answer is available decides whether it gets used. Building for the phone changed which decisions the product could influence, because the ones worth influencing are mostly made away from a desk.

Privacy can be an architecture, not a policy. Keeping data in the browser and making the backend stateless turned the hardest enterprise objection, "where does our data go?", into a one-sentence answer.

AI earns trust through transparency. Showing Glaukos's reasoning, and framing outputs as recommendations to a human decision-maker, made the AI a credible analyst instead of a gimmick.