Capstone deliverable: working prototype, architecture, ROI in PKR
اختتامی deliverable: چلتا prototype، ساخت، اور PKR میں ROI
40 min read
Three ways to see it
Component one: the working prototype. Five real (or synthetic) inputs flowing through the chosen architecture, producing five outputs that an actual user has verified as 'directionally right'. The prototype lives on the engineer's laptop or a private staging environment, not yet production. The README explains in one page how to run it, with synthetic test inputs included. The prototype need not be beautiful. It must be reproducible. A reproducible prototype is something anyone in your team can run; a beautiful one is sometimes only the engineer who built it.
Component two: the architecture document. Two pages. Page one: the four-box flowchart (trigger, middleware, model/RAG/agent, system of record) with the data each arrow carries, the authentication on each boundary, and one error-handling note per arrow. Page two: the data classification table, the chosen deployment option, the eval-set summary with current metrics, the monitoring runbook in three lines, the five risks each with one mitigation. If the document is longer than two pages, your scope is too wide. The capstone test is whether you can compress everything you have learned into two pages.
Component three: the ROI calculation in PKR. Three numbers. Number one, current monthly cost of the manual process being replaced, in PKR person-hours times salary plus overhead. Number two, monthly cost of the AI system at projected steady-state volume, in PKR including model, infrastructure, and amortised build effort. Number three, the difference, expressed as months until payback and as a per-year saving once payback is reached. Add a one-paragraph honesty section: where the savings might fail to materialise, what assumptions are most fragile, what the worst-case looks like. The honesty paragraph is what a CFO trusts. The promise paragraph is what a salesperson writes.
Quick check
Quick check: what makes modern AI different from a rule-based program?
The why-tree
Why-tree level one: why all three components and not just the working prototype? Because a prototype without architecture documentation cannot be handed off. A prototype with architecture but no ROI cannot be funded. A prototype with ROI but no architecture cannot be defended. All three together are what take a pilot to a programme.
Try this with Claude
AI-edge prompt to try with Claude or ChatGPT: 'Critique my capstone, attached. Output three sections: weakest claim with one line on how to strengthen it, biggest hidden cost with one line on how to disclose it, single most likely failure mode in production with one line of mitigation. Then write the six-line executive summary I will read aloud to my board, half English half Urdu.' Iterate.
Sources
Sources and further reading. Anthropic, 'Build with Claude' (docs.anthropic.com). 'Software Engineering at Google' chapters on launch and retrospectives (abseil.io/resources/swe-book). McKinsey, 'The state of AI in 2025' for benchmarking ROI claims. Andrew Ng's 'A Career in AI' guide (deeplearning.ai). Pakistan Personal Data Protection Act 2023 (na.gov.pk). Throughout this track: Anthropic, OpenAI, Google AI docs; MCP specification at modelcontextprotocol.io; MDN HTTP and API docs; Postman Learning Center; OpenAPI 3.0 spec.