Job search
FinOps Portfolio for Applications: Explain a Cloud Bill and Test a Cost Recommendation
Separate allocation, future spending assumptions and a tested model recommendation, retaining unallocated costs, credits and capacity failures as reviewable evidence.
A FinOps portfolio should explain where a bill’s money went and show whether one cost recommendation survives an engineering constraint. Keep allocation, forecasting and optimization as separate deliverables. Reassigning a shared cost does not reduce the invoice; a cheaper capacity model does not establish real savings or acceptable application performance.
This guide actually runs two original local exercises: Decimal arithmetic over a synthetic bill, and a deterministic virtual-time queue model testing a worker-capacity recommendation. Every price, usage quantity and workload is invented. No cloud account, real invoice, spending action or resource deletion was involved. The queue experiment is not a CPU benchmark or live load test. LandOffer publishes this guide.

Use recent roles on LandOffer.ai to identify the application-cost responsibility your work sample should demonstrate, then provide the evidence needed to inspect your reasoning.
Separate the three questions your portfolio answers
The FinOps Foundation’s current Allocation capability describes assigning direct and shared costs through accounts, tags and other metadata. It calls for allocation, tagging and shared-cost strategies. That supports explaining ownership choices rather than silently assigning every unknown charge to the largest application.
Its Forecasting capability concerns a model of future spending for a defined scope, informed by historical spending and planned changes. A forecast can inform a budget; the two are distinct capabilities. Label the reconciled bill as historical sample data and the projection as a conditional scenario, with its assumptions beside it.
The current optimization capability is Usage Optimization, previously called Workload Optimization. Its guidance connects resource use with functional and nonfunctional requirements and trade-offs among cost, performance and risk. The portfolio task is therefore to test a recommendation against a stated constraint, not just identify a lower number.
These sources provide the framework; the following datasets, policy choices and results are our own synthetic work. We did not reproduce an organization’s allocation policy, establish fair chargeback or validate provider prices. Current public documentation does not supply those missing facts for a particular employer.
Completed asset one: reconcile the synthetic bill
The fictional thirty-day bill uses USD teaching amounts. It is not an AWS, Azure or Google Cloud quote, and its simple rates do not represent taxes, commitments or provider billing rules. The script verifies each quantity-times-rate calculation before assigning costs.
| Synthetic line | Quantity and fictional rate | Amount | Assignment policy |
|---|---|---|---|
| Iris compute | 720 worker-hours × $0.10 | $72 | Direct to Iris |
| Iris storage | 200 GB-month × $0.02 | $4 | Direct to Iris |
| Beacon compute | 360 worker-hours × $0.10 | $36 | Direct to Beacon |
| Beacon storage | 100 GB-month × $0.02 | $2 | Direct to Beacon |
| Shared network | 500 GB × $0.04 | $20 | Requests proxy: Iris 60%, Beacon 40% |
| Shared observability | 100 GB × $0.10 | $10 | Log-volume proxy: Iris 80%, Beacon 20% |
| Unowned log charge | 50 GB × $0.10 | $5 | Leave unallocated pending ownership evidence |
| Support | One fixed-period charge | $10 | Hold centrally under the stated policy |
| Iris-specific credit | One nonrecurring credit | −$6 | Apply only to Iris |
Positive charges total $159. The credit is −$6, producing a $153 net bill. Those are different totals; dropping the credit or quietly subtracting it from a shared pool would change the application-level explanation.
The completed allocation is Iris $90, Beacon $48, central $10 and unallocated $5. Iris receives $76 direct costs, $12 network and $8 observability, then its $6 credit. Beacon receives $38 direct costs, $8 network and $2 observability. The four outputs sum exactly to $153.
The result preserves a useful unresolved item. Retain the $5 unowned charge in the reconciliation and assign an ownership question rather than deleting the row. Its owner is unknown in the authored data. A reviewer can see what requires investigation without mistaking a balancing spreadsheet for complete ownership evidence.
The two shared drivers are also deliberate assumptions. Requests may be a proxy for network use; log volume may be a proxy for observability consumption. Neither is proven fair by the arithmetic. Record who would approve each policy in a real organization and what evidence could justify changing it.
Explain a change without calling it waste
The sample includes a prior network quantity of 250 GB at the same fictional $0.04 per GB, costing $10. The current 500 GB costs $20. The $10 increase comes from doubled quantity in this model, not a rate increase. It does not prove that either period contained waste or that the network service malfunctioned.
Likewise, a larger application allocation can reflect a changed shared-cost driver rather than additional total spending. If finance changes how the same $20 pool is divided, the application totals change while the invoice stays constant. An allocation improvement is valuable transparency, but it is not a savings event.
The project also declares completed-report denominators: 12,000 for Iris and 8,000 for Beacon in the same synthetic period. Their allocated costs per report are $90 / 12,000 = $0.0075 and $48 / 8,000 = $0.006. These include the stated shared allocations and Iris credit, while excluding central and unallocated costs.
Do not rank the applications as equally useful products from those two unit costs. The exercise does not establish comparable report complexity, customer value or service quality. Changing the denominator to attempted requests would answer a different question. Keep units, period and cost scope beside the figure.
Forecast a change as a separate scenario
Our conditional next-period scenario raises network quantity by 25%, from 500 to 625 GB, and removes the nonrecurring $6 credit. It holds every other quantity and rate fixed. Network cost becomes $25, and the modeled net total becomes $153 + $5 + $6 = $164.
That is a completed arithmetic scenario, not an approved budget or a validated forecast. It assumes that other costs do not move with demand, which may be unrealistic for an actual application. There is no historical series, demand confidence interval or business approval supporting those assumptions here.
A reviewer should be able to change the growth assumption and rerun the calculation without editing a claimed actual bill. Preserve the original allocation result and place the scenario in a separate output field. Otherwise, a forward-looking estimate can overwrite the evidence needed to explain the past period.
The result also shows why credits deserve explicit treatment. Projecting the $153 net amount unchanged would retain a benefit that this scenario says expires. A real forecast would need the applicable credit terms and remaining balance, current rates, workload changes and the responsible owners’ agreement.

Completed asset two: test a cheaper capacity plan
The independent experiment is a discrete-event simulation of a fictional batch worker queue. Python actually computes assignments, starts, finishes and cost arithmetic; the durations are virtual model inputs, not measured application execution times. This experiment does not resize a real instance or run work against a cloud service.
The one-hour window contains 32 jobs: eight arrive at second zero, sixteen at second 900 and eight at second 1800. Each requires exactly 90 virtual seconds on one worker. Jobs are assigned in arrival order to the earliest available worker. The scheduler rejects a worker interval if the full 90-second job would extend beyond it; jobs are not interrupted or silently carried into an unreserved window.
The predeclared guardrail is completion within 900 seconds of each job’s arrival, with all 32 jobs completed. The model’s fictional price is $0.12 per reserved worker-hour, with linear billing and no minimum rounding, taxes, discounts or startup charges. It bills the reserved window, not just busy processing time.
| Plan actually evaluated | Worker-hours | Modeled cost | Maximum arrival-to-finish time | Jobs over guardrail |
|---|---|---|---|---|
| Two workers available for the full hour | 2 | $0.24 | 720 seconds | 0 |
| One worker for the full hour | 1 | $0.12 | 1,440 seconds | 10 |
| One all hour; second during seconds 900–1800 | 1.25 | $0.15 | 720 seconds | 0 |
| Same billed peak window, but second worker unavailable for its first ten minutes | 1.25 | $0.15 | 1,170 seconds | 4 |
All four runs complete 32 jobs within the one-hour window. That alone is insufficient: ten jobs in the one-worker plan miss the stricter arrival-to-finish requirement. Checking only whether all jobs finish during the hour would accept the one-worker plan despite ten missed per-job deadlines.
The peak-only second worker uses 1.25 worker-hours rather than two. Its modeled difference is $0.09, or 37.5% of the $0.24 baseline, under the explicitly fictional rate and billing assumptions. This is not money saved on an invoice. The model has no provider meter, bill, deployment cost or operational labor.
Retain the failed sensitivity test in the decision
The fourth run changes one assumption: the second worker is billed from second 900 but cannot process jobs until second 1500. Its ten-minute warmup consumes part of the reserved interval. The resulting four late jobs show that the favorable peak-window result depends on immediately usable capacity.
The completed decision is to reject the one-worker plan and withhold real rollout of the peak schedule pending authorized measurement of readiness and workload behavior. Retaining the two-worker baseline is the model review’s decision. It does not establish permission or capacity to restore a real service, and no system was changed. A candidate should not present the peak plan as approved merely because its optimistic case passed.
The paper rollback boundary is explicit: if usable capacity arrives late or the agreed completion threshold is breached during a future authorized pilot, stop the reduced-capacity experiment and restore the approved baseline through its owner. No rollback command was executed, and this article does not supply authority to change resources.
The script preserves each job’s arrival, selected worker, start, finish and arrival-to-finish duration. It checks that completed service time totals 32 × 90 = 2,880 worker-seconds, and that jobs never overlap on the same worker. These checks validate finite model accounting; they do not prove production reliability.
A real recommendation would also need observed job-duration variability, retries, memory and I/O limits, scheduling overhead, resilience requirements and true billing behavior. An observed CPU average would not settle all of those questions. Keep the missing evidence visible rather than calling the simulated guardrail a measured application SLO.
Package the work so someone can inspect the result
The local project includes the bill CSV, allocation-policy JSON, job arrivals, capacity plans, standard-library Python runner and event-level outputs. It records Python 3.12.14, input hashes and the exact run command. From the project root, the recorded executable runs the report’s run_lab.py; the README contains the complete project-relative path.
Two actual fresh-process executions produced identical saved result hashes; the record identifies the tested script and inputs. There is no cloud SDK, credential, account number or private cost export. The script regenerates its deterministic original fixtures, so adding another scenario requires an intentional edit to the authored inputs rather than silently replacing an employer’s data file.
The portfolio statement can be precise: “Reconciled a synthetic net bill with explicit shared-cost drivers and unresolved ownership, then executed a deterministic queue model that rejected a cheaper capacity plan and exposed a warmup assumption in another.” Keep the 37.5% figure attached to this model’s fictional worker-cost comparison; no company spend or realized savings was measured.
Use that evidence when you browse recent roles on LandOffer.ai. The demonstrated contribution is reproducible analysis and a bounded engineering decision. Real allocation approval, workload performance, deployment and realized savings remain separate evidence to obtain through authorized work.
Download the worked example files
Download the curated educational companion. Read PUBLICATION-README.md after extraction for the public file layout, replay commands and requirements. The listed offline Python checks were rerun from the extracted archive in a disposable directory. The earlier local report and commands describe the original evidence environment; use the companion README for the portable layout. No whole evidence directory, credentials or compiler binaries are included.


