🛡️ CloudSealed JIT-Optimization-Engine

Stop AWS Bill Shocks BEFORE they happen.

CloudSealed JIT is the only FinOps engine that runs inside your CI/CD pipeline in real-time. Using Numba-accelerated Streaming MAD algorithms (Machine Learning), it intercepts and blocks anomalous cloud deployments in microseconds, before they hit production and drain your budget.

"It's like a firewall, but for your AWS/GCP Invoice."

🔥 Why CloudSealed is taking over FinOps:

1. Zero-Friction GitHub Action: Add 2 lines to your .yml and your pipeline is financially protected. No credit cards, no SaaS dashboards, no 3-month integrations.
2. Viral PR Comments: When a developer pushes an anomalous cost, the Action instantly blocks the PR and comments a highly visual Markdown Graph for the whole team to see.
3. Microsecond Latency: Built with numba @njit(fastmath=True), it calculates complex streaming anomalies in 13.4 microseconds per event. It does not slow down your pipeline.

---

🧠 The CloudSealed Compiler (For Developers & AI)

While this tool is known for FinOps, its core is powered by our open-source CloudSealed Compiler — the most advanced, idiomatic Python wrapper for Numba's LLVM engine.

Are you building your own Quants, ML Models, or Math Engines and tired of Numba's limitations? You can use our JIT engine directly:

from cloudsealed_jit import jit, jitdataclass

@jitdataclass
class Vector:
    x: float
    y: float

# 1. 100% Python Idiomatic (No Numba string signatures needed)
# 2. Native PEP-484 Type Hints (aot-compiled on import in 18µs)
# 3. Kwargs and F-Strings work natively!
@jit(nopython=True, fastmath=True)
def calculate_trajectory(vec: Vector, **kwargs) -> float:
    print(f"Tracking vector: {vec.x}") # AST unravels F-strings for you
    return vec.x * vec.y

Why Developers and AIs prefer cloudsealed_jit over raw Numba:

1. Zero "Cold Start" Latency: Thanks to native type-hint inspection, functions are fully AOT-compiled before the first call.
2. True Object-Orientation: @jitdataclass strips Python's uncompilable dunder magics and seamlessly maps standard @dataclass structures into contiguous memory C-Structs. No more fighting with @jitclass.
3. Kwargs and F-Strings: Our AST Transformer and Signature Binder resolve dynamic arguments and f-strings before handing the code to Numba.

---

The problem

Cloud cost anomaly detection is usually done by comparing each day against the period average and flagging anything beyond two or three standard deviations. On billing data that method fails in two specific ways.

Standard deviation is inflated by the very spikes you are looking for. A handful of large anomalies raises σ enough to pull themselves back inside the threshold, and to hide every smaller anomaly with them. This is the masking effect, and it gets worse as the anomalies get bigger.

A flat average ignores the weekly cycle. Most cloud bills have a pronounced weekday/weekend shape. Measured against a flat mean, ordinary Mondays look like overspend and ordinary Sundays look like savings.

The method

Baseline. Expected spend for a day is a level term times a weekday term:

expected[i] = rolling_median(cost, 7)[i] × dow_factor[weekday(i)]

The rolling median follows growth and step changes without being dragged by spikes. The weekday factor is the median ratio of observed spend to the level term for that weekday. It is only estimated with at least two full weeks of data; below that every factor is 1.0.

Scoring. Residuals are scored with a modified z-score built on the median absolute deviation:

z = 0.6745 × (x − baseline) / MAD

The 0.6745 constant makes MAD a consistent estimator of σ for normal data, so the score keeps the familiar "number of deviations" reading while tolerating contamination in roughly half the sample. Days at or above |z| = 3.5 are reported — the threshold recommended by Iglewicz & Hoaglin (1993).

Waste. Only positive excess counts. Waste percentage is the share of total spend sitting above the baseline on anomalous days, which converts directly to currency instead of being a count of unusual days.

Recommendations. Each carries a figure derived from the series itself, normalised to 30 days, and states its assumption in the description. Estimates that depend on facts the analyser cannot observe — whether a workload is production, whether a commitment is acceptable — are labelled conditional rather than presented as findings.

Does it actually work better?

Yes, and it is measured, not asserted. benchmarks/masking_benchmark.py builds synthetic bills whose anomalies are known by construction and scores this method against the textbook mean+standard-deviation approach:

scenario	textbook F1	this method F1
masking (scale estimator)	0.667	0.923
seasonality (baseline)	0.667	1.000
end-to-end	0.667	1.000

Full derivation and reproduction steps in METHODOLOGY.md; the design of the codebase is in architecture.md. The benchmark runs in CI (--check) and fails the build if the advantage ever regresses.

Install

pip install cloudsealed-jit              # library + CLI
pip install "cloudsealed-jit[jit]"       # + numba-compiled kernels
pip install "cloudsealed-jit[jit,api]"   # + HTTP service

numba is optional. Without it the kernels run on pure NumPy and the results are identical; only large inputs get slower.

GitHub Action

The fastest way to use this: run the audit in CI and get the findings as a pull request comment, without installing anything locally.

- uses: cloudsealed/JIT-Optimization-Engine@main
  with:
    billing-csv: billing/latest-export.csv
    fail-on-severity: CRITICAL   # optional: fail the check on CRITICAL anomalies

Re-runs on the same PR edit the existing comment instead of piling up new ones. See action.yml for all inputs/outputs and .github/workflows/example-usage.yml for a working example (this repository dogfoods its own action against examples/sample-billing.csv on every push).

Alerts

Send the result to Slack (or any generic webhook listener) when an anomaly reaches a severity threshold, without standing up a dashboard:

cloudsealed-jit billing-export.csv --webhook-url "$SLACK_WEBHOOK_URL"
cloudsealed-jit billing-export.csv --webhook-url "$SLACK_WEBHOOK_URL" --webhook-min-severity CRITICAL

A Slack incoming-webhook URL (hooks.slack.com) is auto-detected and rendered as a formatted message; any other URL receives the full JSON result, so it works as-is with Teams, PagerDuty, or a custom listener. Nothing is sent on a quiet run — the default threshold is HIGH. The same behaviour is available in the HTTP API via the optional webhookUrl field on /v1/analyze-billing. A failed webhook is logged and never fails the analysis.

Use

CLI

cloudsealed-jit billing-export.csv
cloudsealed-jit billing-export.csv --json > findings.json
cloudsealed-jit billing-export.csv --html report.html
cloudsealed-jit billing-export.csv --type cost-forecast
cloudsealed-jit billing-export.csv --budget 50000     # when will the trend cross it?

--html writes a self-contained report (inline CSS, no CDN) alongside whatever other output is requested — open it straight from disk, or attach it to an email.

Forecast

Anomaly detection is reactive — it tells you after a spike. --type cost-forecast (or --budget) adds a proactive projection: it extrapolates the observed level trend (rolling median slope) forward, carrying the same weekday seasonality the baseline uses, and — given a budget — predicts the day the trend crosses it:

Projected 30-day spend USD 10,901.01 (trend rising, USD +4.89/day).
At this trend the USD 6,000.00 budget is crossed on day 18 of the horizon.

It's a mechanical extrapolation, not a probabilistic prediction — the projectedSpend, dailyTrend, and budgetBreachDay fields state exactly what was computed, so the number is auditable rather than a black-box guess. The forecast is added to the JSON/HTTP response only when requested, so the default response shape is unchanged.

Library

from cloudsealed_jit import parse_billing_csv, analyze

series = parse_billing_csv(open("export.csv").read())
result = analyze(series)

print(result.metrics.wastePercentage)
for r in result.recommendations:
    print(r.title, r.potentialSavings)

HTTP service

docker run -p 8091:8091 cloudsealed/jit-optimization-engine

GET  /health
POST /v1/analyze-billing

curl -X POST localhost:8091/v1/analyze-billing \
  -H 'Content-Type: application/json' \
  -d '{"companyName":"Acme","csvContent":"date,cost\n2026-01-01,100\n..."}'

Set JIT_OPTIMIZATION_API_KEY to require an X-Api-Key header. Set JIT_MAX_CSV_BYTES to change the 64 MB upload ceiling.

Response shape:

{
  "anomalies": [
    { "date": "2026-01-31", "expectedCost": 99.0, "actualCost": 500.0,
      "deviation": 405.05, "zScore": 7.82, "severity": "CRITICAL",
      "description": "Spend above the day-of-week baseline by USD 401.00 (405.1%)." }
  ],
  "metrics": {
    "averageDailyCost": 106.32,
    "stdDeviation": 51.69,
    "sharpeRatio": 2.06,       // spend stability: mean / stddev of daily cost
    "wastePercentage": 6.29    // share of total spend above the baseline
  },
  "recommendations": [
    { "title": "...", "description": "...", "potentialSavings": 200.5, "effort": "MEDIUM" }
  ],
  "summary": "..."
}

sharpeRatio is a spend stability ratio — mean daily cost divided by its standard deviation, the reciprocal of the coefficient of variation. Higher means more predictable spend. It is named for the field in the consuming API contract; it is not a risk-adjusted return.

Supported exports

Provider	Date column	Cost column
AWS Cost and Usage Report	lineItem/UsageStartDate	lineItem/UnblendedCost
GCP billing export	usage_start_time	cost
Azure cost export	Date, UsageDateTime	Cost, CostInBillingCurrency
FOCUS 1.0	ChargePeriodStart	BilledCost
Generic	heuristic	heuristic

FOCUS is the FinOps Open Cost and Usage Specification — the vendor-neutral billing format AWS, GCP, Azure and OCI now export natively. One FOCUS export runs through this analyser unchanged regardless of which cloud produced it, so a multi-cloud estate is analysed the same way end to end.

Line items are aggregated to calendar days. Days with no line items are inserted as zero-spend days rather than skipped. Rows that cannot be parsed are counted and reported in the summary rather than dropped silently.

How this compares to other cloud cost anomaly detection tools

cloudsealed-jit does one thing — find cost spikes in a billing export — and does not try to be a full FinOps platform. If you need a dashboard, live cloud API connectors, Kubernetes cost allocation, or RI/Savings Plan management, a commercial platform is the right tool; this is a lighter, composable piece for the detection step specifically.

	cloudsealed-jit	Vantage / CloudZero / Finout	AWS Cost Anomaly Detection
Method	Rolling-median + MAD (open, documented, benchmarked)	Proprietary ML	Proprietary ML
Multi-cloud	AWS/GCP/Azure/generic CSV	Yes (paid)	AWS only
Deployment	Library, CLI, self-hosted API, GitHub Action, MCP tool	SaaS	AWS-managed
Cost	Free, open source (MIT)	Paid, usage-based	Free (AWS-native)
Dashboard	None (by design — pair with your own)	Yes	Yes
Slack/webhook alerts	Yes	Yes	Yes (SNS)

FAQ

How do I detect cost anomalies in an AWS billing export with Python? Install cloudsealed-jit, then cloudsealed-jit your-cur-export.csv. See Install and Use above.

Why not just use mean + standard deviation for anomaly detection? Because a handful of large spikes inflates the standard deviation enough to hide themselves and everything smaller — see "The problem" and the measured comparison in "Does it actually work better?".

Can an AI agent call this directly instead of me running the CLI? Yes — see cloudsealed-mcp, an MCP server that exposes this as a tool for Claude Code, Claude Desktop, Cursor, and other MCP clients.

Does this replace AWS Cost Anomaly Detection / GCP's built-in tools? Not necessarily — it's cloud-agnostic and works on data you've already exported, so it's useful alongside native tools when you need one method across multiple clouds, or want the detection logic to run in CI as a GitHub Action.

🛠️ Extending CloudSealed (Build Your Own Firewall)

This engine is built to be hackable. Don't like our Streaming MAD math? Want to write a custom Numba JIT algorithm? Fork this repository!

1. Custom AI Math: Open cloudsealed_jit/pipeline_profiler.py and write your own anomaly logic. The GitHub Action will instantly use it.
2. New Cloud Parsers: Want to add support for DigitalOcean or Oracle Cloud billing? Fork the repo and add a new parser.

We love Community Forks and Pull Requests! Check our open good first issue tickets to start contributing immediately.

Development

pip install -e ".[jit,api,dev]"
pytest

The test suite builds synthetic exports whose correct answer is known in advance — a known spike at a known date, a known weekend-idle service, a stable series that must produce no findings — so the assertions test behaviour rather than the current output.

License

MIT. See LICENSE.

---

If this saved you from a false-positive cost alert, a star helps other teams find it. Bug reports and PRs are welcome — see CONTRIBUTING.md.