FinOps Reporting & Budget Platform Case Study

Where It Started

The Monthly Spreadsheet Cycle

Actuals were copied into Excel every month, and year-end projections were educated guesses.

100+Cloud subscriptions, each pivoted through a provider billing console to find variance

16Solutions across the organization, with a taxonomy that matched nobody else’s naming

3Providers, and no single view of where the money was going or where it was headed

Leadership needed answers to three questions. Are we going to hit budget this year? Which subscriptions are driving overruns? What changed month over month?

And they needed them without waiting for someone to update a spreadsheet. That turned the project into a data pipeline first and a dashboard second.

The Pipeline

From Export To Dashboard

Left to right, with no manual export anywhere in the path.

Ternary API

FOCUS billing feed for AWS, Azure and GCP, up to 50,000 rows a day. Azure uses EffectiveCost (amortized); AWS and GCP use BilledCost.

Normalize

Three alias maps translate provider names, subscription codenames and business unit codes. Records are de-duplicated and upserted.

PostgreSQL

Two databases: core financials (cloud_budget) and detailed vendor spend (vendor_spend).

Express And React

One Express server serves the API and the React app from the same container: dashboard, scorecard, insights and report pages.

It runs on startup. syncCurrentMonth() fetches current-month actuals, extrapolates a month-to-date forecast and projects the remaining fiscal year months. No scheduled task is needed.
History loads once. A backfill script populates FY23 (July 2022) through the present from historical billing, and a marker file stops it running twice.
Gaps are filled on purpose. Fiscal year gap-fill creates empty records for all 12 months per subscription, so the budget entry page always has complete period coverage.
A new alias is one line of config and a redeploy, not a data migration.
The accepted risk: if Ternary is down at startup, the app serves stale actuals. Given a daily reporting cadence, that beat manual exports.

Two competing budget figures in one tool is worse than one manual import step.
Ternary has a budget feature, and syncing it would have been the easy call. I turned it off. Its budget entries are scoped too inconsistently to reconcile against the finance-owned operations workbook, which is the number leadership actually governs against. Budgets now flow from that workbook through one import page, and every budget-derived view reads from that single source.

The Grading Rules

What Goes Into A Grade

Each solution gets a letter from A+ through D−, built from five weighted dimensions.

Dimension	Weight	How It Is Measured
Budget Adherence	45%	Actual against budget, asymmetric by design (see below)
Risk & Anomalies	20%	Anomaly and risk signals; the insights tabs flag spikes over 25% month over month and dormant spend under $1K a month across 3 or more months
Spend Stability	15%	How steady monthly spend is, so volatile subscriptions grade lower
Cost Efficiency	10%	Computed in code from the same actuals
Forecast Accuracy	10%	Computed in code from the same actuals

Same Percentage, Different Grade

Case	Budget	Overrun	Percent Over	Relative Grade
Small subscription	$10K	$5K	50%	Better
Large subscription	$1M	$500K	50%	Worse

Over budget costs twice as much as under. A naive distance-from-budget score treats 10% under and 10% over as equally bad. Overspending is far more damaging, so it is penalized at twice the rate.
Magnitude counts, not just percentage. The penalty scales with absolute dollar variance, which is why the $500K overrun scores worse than the $5K one.
Every component is visible. The UI shows each part of every grade, so stakeholders can see exactly why a subscription received its score and the design nudges decisions toward the right priorities without judgment calls on each one.

Looking Ahead

Forecasts And Flags

Year-end projections stopped being guesses once each horizon had its own method.

Horizon	Method
Current Month	Extrapolated from the month-to-date daily average
Future Months	The current daily rate, adjusted for days in the month
Rolling And Year-End	Linear regression on the last 6 months of actuals
Budget Pre-Fill	Last actuals, a 3-month average, or a linear trend, chosen by the admin

Budget alerts have four states: Critical, At Risk, Warning and On Track.
Movement is configurable. Month-over-month and year-over-year changes use thresholds an analyst can adjust, plus a COGS and non-COGS view for gross margin reporting.
The executive page follows the FinOps Framework. Headline KPIs (annual budget, projected year-end, budget remaining, active subscriptions) sit above one red, amber or green card per domain (Understand, Quantify, Optimize, Manage), each with a one-line verdict that scrolls to the page proving the number. Headline data loads in one call, and heavier detail loads lazily so first paint is never blocked.

Design Decision

The Fiscal Year Is The Calendar

Every component thinks in fiscal years, July 1 through June 30. That was settled before the first line of code.

Component	Fiscal Year Behavior
Date Pickers	Default to fiscal year ranges
Scorecards	Filter to the current fiscal year automatically
Forecasts	Project through June of the current fiscal year
Budget Entry	All 12 fiscal months, grouped by quarter
July 1	Everything rolls to the new fiscal year with no manual configuration

A calendar-year system would have needed correcting every July. Fiscal-native design meant the app matched how leadership already thought about budgets from day one.

Built On The Same Data

One Dataset, Five Outputs

Once the pipeline was trusted, everything else became a view over it.

Output	What It Does
Monthly Report	Renders the Monthly Cloud Report’s chart and table pages straight from the database, sized for the executive deck, and reconciles to the published deck figures line for line. It replaced charts that were rebuilt by hand in Excel every month.
Vendor Tracking	11 vendors with monthly budget versus actual, COGS and non-COGS splits and cumulative variance. Three have solution-level detail, up to 34 solutions and 9 sub-subscriptions in one vendor.
AI Token Spend	Service-level AI spend by business unit next to disaster recovery spend. Two BigQuery-backed views attribute AI coding tool cost to the individual user, with an explicit “unattributed” row so totals always tie.
Analysis Tabs	Eight tabs of analysis, including month-over-month and year-over-year movement, anomaly detection and budget burn-down.
ARIA	Twelve tools in ARIA, the FinOps AI agent I built on top, read from these APIs and answer in plain English with a link back to the exact filtered view.

The query itself costs money. Every BigQuery scan is cached for ten minutes and bounded to a trailing window, and caches invalidate after each AI spend sync.

Trust And Access

Who Can Change What

Roles are enforced in the backend on every write. The frontend only mirrors them.

Role	Can Do
Read-Only	View all dashboards, analysis and reports
Tenant	Read access, plus forecast editing for their own solution area
Admin	Full create, read, update and delete: budget entry, data imports, user management and Clean Slate with Backfill

The API rejects unauthorized requests whatever the client sends. Role checks live in Express middleware on every write endpoint, and write endpoints are limited to 100 requests per 15 minutes.
Two layers of authentication. Okta SSO and VPN-only access at the platform level, then JWT tokens inside the app with a 24-hour expiry, a fresh role lookup on each request and bcrypt password hashing at cost factor 10.
Queries are parameterized throughout, with no string interpolation anywhere in the codebase, and the API key lives in GCP Secret Manager rather than in code or environment variables.

Stack And Trade-Offs

Why Two Databases And A Grid

Two structural choices shaped the app more than any framework did.

Vendor spend lives apart from core financials. Its granularity (34 solutions, 9 sub-subscriptions) does not map cleanly to the subscription schema, it updates manually on a different schedule than the automated actuals, and a vendor restore never touches core records.
Budget entry runs on AG-Grid. A budget for 100+ subscriptions across 12 months is table-heavy work, and stakeholders could click a cell and type, just as in the spreadsheet they knew. It adds sorting, filtering and pinning, bulk apply across a filtered set, and CSV and Excel export.
The Community license has limits, including server-side row models and advanced pivoting. Neither was needed here, so the trade held.
Tests: 129 in total, 80 on the backend and 49 on the frontend, run with Vitest.

React 19 Vite Material-UI 7 AG-Grid Recharts Node.js 20 Express 4 PostgreSQL 15 Vitest Docker Kubernetes ArgoCD

Retrospective

Four Things I Would Change

Each one is a lesson about what becomes expensive later.

Use TypeScript Throughout

The backend especially would benefit from stronger type safety across the normalization pipeline, where a misnamed field is a wrong number.
Log Who Changed What From Day One

Knowing who changed which budget figure, and when, became a frequent question. Audit logging added late cannot recover the history that came before it.
Build Backup And Recovery Early

Database backup and disaster recovery tooling arrived as an afterthought. Dual-database seeding also needed careful transaction management to keep both databases consistent on startup.
Deploy To Kubernetes From The Start

The early Docker Compose setup matched production for development, but moving to GKE sooner would have bought scalability and observability earlier.

The hardest part of FinOps tooling is not the technology. It is the data normalization.
A reliable pipeline that handles different billing models, internal codenames and a taxonomy that matches nobody else’s, and makes new mappings easy when names change, was the foundation everything else depended on. The scorecard taught me to think about what “good” means in financial operations: a plain adherence percentage cannot tell a small subscription missing by 50% from a large one missing by 5%.

Back to Projects

FinOps Reporting & Budget Platform