RevenueGuard

An agentic revenue-leak auditor for B2B SaaS finance teams — it verifies your data is fresh before it trusts it.

🚀 Live demo: https://revenueguard-273367103135.us-central1.run.app — pick the revenueguard app and ask: "Audit my revenue for leaks across all five leak families." 🎬 Pitch video: deploy/revenueguard_pitch.mp4 (YouTube link coming with the submission)

---

Problem

B2B SaaS companies leak 2–5% of revenue to billing errors they never see — failed payments that were never retried, expired cards on active subscriptions, usage billed against cancelled plans, renewals that silently never invoiced, discounts that outlived their expiry. Worse: most leak-detection tools audit stale data, so they cry wolf. Finance and RevOps teams stop trusting the alerts, and the real leaks keep flowing.

Solution

RevenueGuard is an autonomous agent that runs a strict verify → audit → report loop:

1. Verify before audit (the wedge). Before auditing any revenue stream, the agent calls the Fivetran MCP to confirm the relevant connector synced recently and is healthy. If a source is stale or broken, it stops, reports it, and — with your confirmation — triggers a re-sync. No more false alerts from stale data.
2. Audit. Once freshness is confirmed, it runs multi-step leak detection over the landed data in BigQuery across five leak families.
3. Act & report. It produces a ranked report with $ exposure per finding, the SQL evidence, and a recommended action — and keeps the human in control (it proposes; you approve any write).

Architecture

The mandatory sequence the agent's instruction enforces:

1. ① Freshness gate — the agent calls the Fivetran MCP (list_connections, get_connection_details) to check the billing connector. Paused, broken, or last-synced > 24h ago ⇒ it refuses to audit, reports why, and — with your approval — resumes the connector (modify_connection) and triggers a sync (sync_connection).
2. ② Audit — five deterministic leak-family SQL queries run over the Fivetran-shaped tables in BigQuery (bigquery_run_audit, with get_leak_query exposing the SQL evidence on demand).
3. ③ Report — compute_exposure and build_report are plain deterministic functions (no LLM inside): same findings, same dollars, every run.

Text version of the diagram

                    ┌──────────────────────────────────────┐
   Web UI           │           RevenueGuard Agent          │
   (ADK on          │            (ADK LlmAgent)             │
    Cloud Run)      │        model = gemini-2.5-pro         │
        │           │                                       │
        └──────────▶│  Tools:                               │
                    │   1) Fivetran MCP (McpToolset)  ◀── pipeline freshness/health/resync
                    │   2) bigquery_run_audit         ◀── leak SQL over landed data
                    │   3) get_leak_query             ◀── SQL evidence on demand
                    │   4) compute_exposure           ◀── $ per finding
                    │   5) build_report               ◀── ranked markdown/JSON report
                    └───────────────┬───────────────────────┘
          ┌─────────────────────────┴───────────────────────┐
          ▼                                                   ▼
   Fivetran MCP server (uvx, stdio)                  BigQuery dataset
   list/details/modify/sync_connection               revenueguard_demo.*
          │                                                   ▲
          ▼                                                   │
   Fivetran pipeline: stripe_billing connector ──── sync ─────┘

Tech stack

• Gemini 2.5 Pro (via Vertex AI) — the reasoning engine.
• Google Cloud Agent Builder / ADK (google-adk) — the code-first agent surface.
• Fivetran MCP (github.com/fivetran/fivetran-mcp) — pipeline freshness, health, and re-sync, launched over stdio via uvx.
• BigQuery — the warehouse where Fivetran lands the billing data.
• Cloud Run — the public hosted URL.

How we used Fivetran

Fivetran is load-bearing, not bolted on. The agent's freshness gate puts Fivetran MCP calls on the critical path of every audit: it lists the relevant connection, checks its last sync time, and refuses to audit until the data is confirmed fresh — triggering a re-sync when it isn't. Auditing stale data is the #1 source of false revenue-leak alerts, so this gate is the whole reason the audit can be trusted.

fivetran_toolset = McpToolset(
    connection_params=StdioConnectionParams(
        server_params=StdioServerParameters(
            command="uvx",                                   # isolates the server's deps
            args=["--from", str(_MCP_DIR), "fivetran-mcp"],  # vendored clone
            env={
                "FIVETRAN_API_KEY": os.environ["FIVETRAN_API_KEY"],
                "FIVETRAN_API_SECRET": os.environ["FIVETRAN_API_SECRET"],
                "FIVETRAN_ALLOW_WRITES": "true",             # enables resume + re-sync
            },
        ),
        timeout=60,  # live API calls exceed ADK's ~5s default
    ),
    tool_filter=["list_connections", "get_connection_details", "get_connection_state",
                 "run_connection_setup_tests", "modify_connection",
                 "sync_connection", "resync_connection"],
)

The gate detects a paused/stale connector, the agent proposes a re-sync, and on your approval it resumes the connector (modify_connection) and triggers a sync (sync_connection) before it trusts a single row.

Screenshot of the Fivetran dashboard showing the connections goes here.

The five leak families

Leak family	What it catches
failed_payments	Failed charges never retried (involuntary churn)
expired_cards	Expired cards on still-active subscriptions
usage_on_cancelled	Active usage billed against a cancelled plan
overdue_renewals	Renewals past due with no invoice raised
expired_discounts	Discounts still applied months past expiry

The seeded demo plants a fixed set of leaks totalling a memorable $18,650/yr in recoverable revenue.

Setup

pip install .                                   # 1. install deps (needs uv/uvx for the MCP)
cp .env.example .env                            # 2. fill in GCP + Fivetran creds
git clone https://github.com/fivetran/fivetran-mcp vendor/fivetran-mcp  # 3. vendor the MCP
python scripts/seed_bigquery.py                 # 4. seed the deterministic demo data
adk web                                          # 5. run the agent locally

Then ask: “Audit my revenue for leaks across all five leak families.” Watch it gate on the Fivetran connector, re-sync it, then return the ranked $18,650/yr report — every finding with its $ exposure and SQL evidence.

Requires uv (provides uvx), which launches the Fivetran MCP in an isolated environment. The shipped agent runs on gemini-2.5-pro via Vertex AI; set RG_MODEL to override.

Roadmap

• Promote the single agent to a FreshnessGate → Auditor two-agent sequence.
• Add write-back actions (auto-dunning, back-billing) behind human approval.
• Expand the leak library (proration errors, tax mismatches, FX drift).
• Scheduled audits with Slack/email digests for finance teams.

License

Apache-2.0.