Suite

Suite is an AI-powered Business OS that automates, reviews, syncs and analyzes financial data, detects risks, generates insights, and creates actionable tasks turning business data into decisions.

Comment

## Inspiration

Every small-business owner we know runs the same painful ritual at month-end:
export transactions from the bank, paste them into a spreadsheet, guess at categories, eyeball the numbers, and hope they set aside enough for tax. The data lives in five different places and the "analysis" is manual, late, and error-prone.

We didn't want to build another chatbot that answers questions about finances. We wanted an agent that actually does the month-end review — plans the steps, moves the data, reasons over it, and leaves the owner with a finished report and a concrete to-do list. The hard part isn't the math; it's getting clean, trustworthy data in front of the model. That's exactly where Fivetran became the agent's superpower.

## What it does

Suite is a financial-operations platform for small businesses. Its flagship is the Financial Review Agent, a multi-step mission that runs end to end with the owner in control:

  1. Checks connected accounts (bank links via Plaid).
  2. Triggers a Fivetran sync through the Fivetran MCP server — landing the business's operational MongoDB data into BigQuery.
  3. Verifies the sync completed before trusting the numbers.
  4. Retrieves the latest transactions from the Fivetran-synced BigQuery warehouse.
  5. Categorizes every transaction. 6–8. Computes revenue, expenses, net cash flow, and an estimated tax reserve.
  6. Compares actuals against budget targets.
  7. Detects anomalies — expense spikes vs. the prior period and dangerous
    category concentration.
  8. Generates an executive summary and recommendations with Gemini.
  9. Takes action — writes action items / tasks, not just text.
  10. Saves the report.

Every one of these 13 steps streams live to the dashboard over Server-Sent Events, so the user watches the agent plan and execute in real time — and stays
in the loop the whole way.

## How we built it

  • Google Cloud Agent Builder + Gemini (Vertex AI) are the brain. Gemini turns the computed figures — revenue, expenses, cash flow, anomalies — into a plain-English executive summary and 2–3 specific, prioritized recommendations.
  • Fivetran (the superpower) is the data backbone. The agent calls the Fivetran MCP server to trigger and verify a connector sync, moving transaction documents out of MongoDB and into BigQuery, where the agent can reason over a clean, query-able analytical warehouse instead of brittle ad-hoc exports.
  • BigQuery is the analytical layer. Fivetran's MongoDB connector lands each document in a JSON data column, so we extract fields with JSON_VALUE / JSON_QUERY and respect _fivetran_deleted for correctness.
  • NestJS (API) orchestrates the mission and owns the tool surface the agent acts through (accounts, transactions, budgets, reports, recommendations, tasks, tax estimation).
  • Next.js (web dashboard) renders the live agent run, the resulting report,
    and the generated action items.
  • MongoDB is the operational source of truth; Plaid provides the bank connections; the whole thing is a Turborepo + pnpm monorepo.

## Challenges we ran into

  • Trusting the data before reasoning on it. An agent that reasons over a half-finished sync produces confidently wrong numbers. We made the agent explicitly trigger the Fivetran sync and verify its status as discrete, visible steps before reading a single transaction.
  • Bridging document data and a SQL warehouse. Fivetran lands MongoDB documents as JSON in BigQuery; getting typed, analytics-ready rows out of that meant careful JSON_VALUE/JSON_QUERY extraction and soft-delete handling.
  • Making autonomy observable. A black-box agent is impossible to trust, so we built the whole run around a live SSE step-stream — the user sees each plan step start, run, and finish with a human-readable detail line.
  • Graceful degradation. Each capability (Fivetran, BigQuery, Gemini) checks whether it's configured and falls back cleanly, so the agent keeps finishing the mission even when one integration is unavailable.

## What we learned

  • An agent is only as good as its data pipeline. Fivetran + MCP turned "the
    data is scattered and stale" from the hardest problem into a single, agent-triggerable step — that's what let the rest of the mission be reliable.
  • "Show your work" is a feature, not a nicety. Streaming the 13 steps made the agent feel trustworthy and kept a human in control of a financial workflow.
  • Separating the brain (Gemini) from the hands (typed tools) and the data (Fivetran → BigQuery) kept each part testable and swappable.

## What's next

  • Richer MCP-driven actions (schedule reviews, file reminders, draft vendor-renegotiation emails).
  • Multi-entity / multi-currency consolidation.

  • Forecasting and cash-runway projection on top of the BigQuery warehouse.

    data is scattered and stale" from the hardest problem into a single, agent-triggerable step — that's what let the rest of the mission be reliable.

  • "Show your work" is a feature, not a nicety. Streaming the 13 steps made the agent feel trustworthy and kept a human in control of a financial workflow.

  • Separating the brain (Gemini) from the hands (typed tools) and the data (Fivetran → BigQuery) kept each part testable and swappable.

## What's next

  • Richer MCP-driven actions (schedule reviews, file reminders, draft vendor-renegotiation emails).
  • Multi-entity / multi-currency consolidation.
  • Forecasting and cash-runway projection on top of the BigQuery warehouse.

Built With

Share this project:

Updates