Simara Architecture Studio

Before you build,
make the decisions clear.

Map the work. Design the handoffs. Put a cost and a test beside the proposal.

AI interpretsSoftware enforcesPeople decide

Define the work

Describe a real process.

Use a concrete example. Include what starts the work, what decisions follow, and what must happen at the end.

Project titleWhat happens today, and what should improve?The business outcome

Available systemsRules and exclusions

Preferred target runtimeTarget inference model for cost estimatesOpenRouter model ID. This is separate from Simara's report writer. Choose No AI for software-only work. Changing provider resets the call mapping for review.

Keep the whole method in view

Work through the three sections below. A blank answer remains an open decision. The first proposal gives you a capability map; return here to confirm how its AI and human steps are costed.

Scope the work: four steps

1. Business requirement to separate capabilities

What starts one request?What counts as completed?

The proposal breaks the process into individual tasks with inputs, outputs and owners.

2. Assign AI, company systems and people

What can be reversed if it goes wrong?What are the consequences of a mistake?Who remains accountable for the business outcome?

The four technical questions below determine where AI fits and which controls it needs.

3. Define the limits of the proposed system

Record input limits, quality, access, freshness and response time. For response time, use the time within which 95% of requests must finish. An average is not the same measure.

4. Carry the limits into the proposed scope

What is explicitly outside this project?Write None only if exclusions have actually been considered.Who will accept the completed work?

The report carries the capabilities, exclusions, assumptions, boundaries and acceptance tests together. It remains a proposed scope until agreed.

Size the use case: four steps

1. Establish demand and count the actual work

Who supplied operational demand?Demand sourceUse business records, not the number of examples in a test dataset.Measurement period

After the first proposal, map the resulting AI and human capabilities here. Until then, the budget and ROI remain incomplete.

All AI steps are covered, or the reviewed design needs no AI.

2. Set token budgets for the distribution of requests

The workload table below sets shares, total input, output, calls and retries. Break down each input here, including the largest case.

Typical request

These three parts must add up to 3,000 input tokens in the workload table.

System instructions (tokens)Leave blank if unknown.Retrieved context (tokens)Leave blank if unknown.User input (tokens)Leave blank if unknown.Largest total input per call (tokens)Leave blank if unknown.

Token distribution source

Cache write costs and reuse

Provider behaviour differs. Cached reads are priced separately from initial writes. Enter no cache savings until the route and reuse pattern are understood.

Input share charged as cache writes (0 to 1)Leave blank if unknown.Cache lifetime in secondsLeave blank if unknown.Seconds between reuse of the same cached materialLeave blank if unknown.Cache eligibility and frequency evidence

3. Price the full operation and assess alternatives

Model prices come from the selected target model. Add all retained staff work, including approvals and exception handling. Group activities performed together once.

All retained human work is covered, or none is required by the reviewed design.

Could requests wait for batch processing?

Does the service deadline allow asynchronous processing?Does the data and compliance configuration allow this route?

4. Test the fragile assumptions

The report recalculates double demand and a more demanding input distribution, with cold cache, more repeated calls and more staff effort. It shows when a scenario crosses the budget ceiling.

Business value and ROI: four steps

1. Establish today's measured baseline

Who owns the baseline?Baseline measurement periodWhat is the baseline based on?

An assumed baseline can show a what-if calculation. It leaves the measured-baseline step open and cannot establish savings or investment payback.

2. Compare the future state in the same units

Why is the projected improvement plausible?Include retained staff work and exceptions. A projection is not an observed result.

Address all five value pillars

Explain the expected contribution, or state why a pillar is not relevant. Different kinds of value should remain visible even when they cannot be priced.

Efficiency: less time or expenseTransformation: work now made possibleProductivity: more output from the same teamSolution cost: what it takes to runService levels: speed and reliability

3. Reuse the full recurring cost from sizing

Basis for the financial caseWhat evidence supports converting this gain into cash?Released time is capacity. Actual cash savings need a change in expenditure or a substantiated income change. The realisable share is entered in the cost table.

4. Show payback and sensitivity

The calculator keeps build cost separate. It shows payback only for a positive cash case, and tests both half the demand and half the gain per request. Technical feasibility and investment value remain separate decisions.

Technical feasibility, in four questions

These are the course assessment questions. Answer each for this use case, naming differences between capabilities. Enter “Unknown” where an answer is missing. Studio must retain that gap and show how each answer affects the design.

Next-token prediction

Does this task require probabilistic generation, or does it require precision on specific values?

Classification, summarization, and drafting are probabilistic tasks where the model excels. Extraction of specific authoritative values (account numbers, policy dates, claim amounts) requires verification against the source of truth.

Where design compensates:

Generator-verifier loops
Code-based evals on extracted values
Tool calls to retrieve quantitative data

Your answer: Next-token prediction

Knowledge

Does this task depend on information that is rare, contested, recent, or domain-specific in ways that may not be represented in training data?

If yes, the design must bring the knowledge into the context window. Do not rely on the model to supply it.

Where design compensates:

Retrieval-augmented generation for stable knowledge
Tool calls for live-state data
Flagging uncertainty on contested claims

Your answer: Knowledge

Working memory

Do the inputs fit comfortably in the context window, or does the task require processing inputs that, in aggregate, exceed the window?

Long documents, multi-document tasks, and extended conversations all hit this constraint.

Where design compensates:

Chunking strategies
Progressive context loading
Summarization across turns
Pipeline architecture for inputs exceeding the context limit

Your answer: Working memory

Steerability

Are the instructions specific, concrete, and verifiable?

Abstract or ambiguous instructions, long reasoning chains, and tasks that require precise numerical or logical computation are all places where the model can drift from intent.

Where design compensates:

System prompts with explicit output schemas
Structured outputs
Code execution for numerical precision
Evaluator-optimizer loops

Your answer: Steerability

Ground the sizing and value case

What do we know about today's work?Give current volume, staff time, errors or turnaround time and where the figures came from. A missing baseline means savings remain unproven.How should the sizing figures be treated?Where did the sizing figures come from?Name the source or explain the assumptions. Include how staff review and the number of AI calls were estimated.

Where this would run, and under what rules

These answers decide which products and routes are even available. Leave an answer unknown and Studio keeps it open as a question; it will not assume one for you.

Who would use it?Is there a cloud commitment?Who is served by one deployment?One request, or a conversation?

Which rules govern this work? Tick every rule that applies. A ticked rule removes options before any trade-off is weighed.

Legally privileged materialRules out: Any route where a third party may retain or review the content.
Patient or other protected health informationRules out: Any configuration not covered by a signed data agreement for that exact configuration.
Data protection and residency rules, such as GDPRRules out: Routes that cannot pin processing to the required region.
Government or public-sector accreditationRules out: Everything outside the accredited environment.
The company's own data or approved-supplier policyRules out: Suppliers and products the policy has not approved.

Which sign-on system says who the user is?Name it, or leave it empty. The server, not the model, decides who the user is and what they may see.Which personal details could appear in a request?One per line. Each one must be answered for: it either enters the request with a reason, or it is withheld.

Requirements and evidence

Record what must hold true. A claimed pass or fail requires a source. Simara preserves your statement, but does not independently certify it.

Workload and cost assumptions

Starting numbers are editable planning assumptions, not measurements. Leave unknown optional fields blank. All costs are USD.

Business requests per monthCurrent baseline hourly staff cost (USD)Monthly hosting (USD)Monthly support and operations (USD)Monthly tools, storage and external services (USD)Monthly budget ceiling (USD)Optional. Blank means unknown.Expected successful completion share (0 to 1)Optional. Blank means unknown.Current manual minutes per requestOptional. Blank means unknown.Share of projected gain realisable as cash (0 to 1)One-off implementation cost (USD)Optional. Blank means unknown.

Your workload mix

Each row represents a request size. Shares must total 1. Calls count every model step in that request.

Workload 1Label

Share (0 to 1)Calls per requestInput tokens per callOutput tokens per callCached input share (0 to 1)Extra calls from retries (0 to 1)

Capacity, quotas and context limits

Use measured processing times and account-specific quotas. These estimates cannot establish a response-time percentile or burst capacity.

Peak business requests per minuteMean processing seconds per requestConcurrent processing slotsProvider requests-per-minute allowanceProvider tokens-per-minute allowancePer-call context limit, including output

Who is this report for?

A verified work email starts the report. We send the finished pack to it and keep your brief on record for you.

Your nameWork emailPersonal mailboxes such as gmail cannot receive a report.CompanyOrganisation sizeYour rolePhone (optional)

I would like Simara to contact me about reviewing this architecture.Send me occasional Simara updates. Optional and separate from the report.

One report per email address. Revisions of this report do not count as a new one.