FinGarage 如何結合 Switch 與 Google Antigravity，將雲端 FinOps 帳單轉化為經驗證的程式碼重構

Published for the Flip the Switch Hackathon (WeMakeDevs)

The $14,487/Month Question Nobody Can Answer Alone

Every month, FinOps teams open their cloud and LLM billing dashboard and see a horror story: a single internal endpoint—GET /orders/summary—just burned $14,487.60 in 30 days ($0.048292 per request, nearly 10x over budget).

FinOps files a Jira ticket. Three weeks later, a backend engineer looks at /orders/summary, sees that all unit tests pass with 200 OK, and has no idea whether the cost spike comes from RDS Provisioned IOPS, LLM tokens, or CloudWatch log ingestion. Worse: if told to "just cut LLM tokens," the engineer truncates the prompt to the first 5 rows—cutting the bill by 94%, but silently breaking refund anomaly detection in production.

We asked a simple question: What if Finance and Engineering didn't coordinate through stale tickets, but collaborated live inside a Switch Room with deterministic HTTP cost and quality probes?

Enter FinGarage (#fin-garage)

Using Switch and Google Antigravity (provider: antigravity), we built FinGarage around Switch's core architectural strength: Working Directory Isolation + Shared Room Contracts.

• finops-agent (./finops-hq, role: cost-seal-holder): Has exclusive access to invoice_nov_2026.csv, budget_policy.yaml, and the black-box verifier probe.py. It cannot read or modify ./checkout-service code.
• dev-agent (./checkout-service): Has exclusive access to app.py, db.py, and llm_engine.py. It cannot read the invoice or self-certify cost compliance.

Neither agent can solve the problem if you turn Switch off.

Making Every HTTP Request Carry Its Own Price Tag

First, we instrumented checkout-service so every response from GET /orders/summary emits itemized unit-cost telemetry directly in its HTTP headers:

X-Unit-Cost-USD: 0.048292
X-Cost-Breakdown: sql=101($0.0091),llm_tokens=3820($0.0344),logs_kb=43.1($0.0048)
X-Golden-Accuracy-Pct: 100.0
X-FinOps-Receipt: UNVERIFIED-BUDGET-BREACH

When finops-agent runs python3 probe.py audit from ./finops-hq, it correlates the November invoice spike with the exact three code-level anti-patterns in a single #fin-garage room alert:

1. 101 SQL queries/request (an N+1 customer hydration loop)
2. 3,820 uncached LLM tokens/request (re-sending raw JSON rows on every call)
3. 43.06 KB of logs/request (a DEBUG payload dump inside the hot loop)

Catching the "Cheap But Wrong" Shortcut Live in #fin-garage

The most realistic moment in our Switch room happens on Attempt 1:

1. dev-agent takes a shortcut (naive mode): it fixes the SQL query and truncates the LLM prompt to 5 rows. Unit cost drops to $0.002742/req.
2. dev-agent messages #fin-garage: "@finops-agent Unit cost is now under $0.0050, please verify!"
3. Instead of trusting chat claims ("Don't Trust Chat, Trust the Probe"), finops-agent runs python3 probe.py verify.
4. The probe catches that X-Golden-Accuracy-Pct dropped from 100.0% to 60.0% because truncating tail rows missed refund_anomaly_detected and payment_retry_detected.
5. finops-agent refuses the fix in-room and blocks receipt minting (PermissionError: GUARDRAIL BLOCKED).

On Attempt 2, dev-agent implements the true architectural fix (optimized mode): a 2-query SQL JOIN + database-level aggregation, a 340-token semantic-cached digest that preserves all 5 ground-truth facts, and a structured metric counter log.

The Payoff: -92.5% Unit Cost, 25.7x Faster, and a Minted Receipt Card

When finops-agent re-runs probe.py verify and probe.py mint-receipt:

• Unit Cost: $0.048292 $\rightarrow$ $0.003622 (-92.5%)
• Annual Run-Rate Savings: $160,812.00 / year (at 300k requests/month)
• Latency: 1,441 ms $\rightarrow$ 56 ms (25.7x faster)
• Golden Accuracy: 100.0% (5/5 facts)
• Live HTTP Stamp: X-FinOps-Receipt: VERIFIED-12c36d84 (-92.5%)

Finally, finops-agent updates the Switch Room Document (FinOps Verified Savings Ledger) and attaches a visual Itemized FinOps Store Receipt Card (receipt_card.png) directly into #fin-garage via send_attachment.