Evidence before precision
Unknowns remain visible instead of becoming confident-looking estimates.
Fictional enterprise scenario · Reviewable method · No customer data
From customer signal to a defensible AI infrastructure decision.
Follow one private-RAG opportunity through discovery, first-pass sizing, architecture framing, financial review, and a conditional executive recommendation.
Ranges before a bill of materials
Early sizing exposes sensitivities and bottlenecks; it does not impersonate a benchmark.
Gates before approval
Architecture, security, benchmark, pricing, and finance reviews remain human decisions.
Interactive decision path
Five stages. One evidence record.
Use the stage controls or left and right arrow keys to move through the opportunity.
Stage 1 · Opportunity discovery
Is the problem specific, material, and sponsored?
The discovery record separates the stated workload and control requirements from assumptions that still need customer evidence.
Known
Private RAG over governed internal content with a stated peak and latency objective.
Unknown
Token distribution, concurrency shape, quality threshold, retrieval design, baseline cost, and support model.
Risk
Premature architecture or pricing precision would conceal the evidence gap rather than reduce it.
Primary output
Workload hypothesis, stakeholder map, evidence gaps, and conditional next action.
Next gate
Named sponsor, representative workload owner, and agreement to a validation-planning workshop.
Executive handoff
Advance conditionally—not confidently.
The use case is specific enough to justify a controlled PoC, but final infrastructure and commercial decisions remain gated by representative workload data, benchmark results, security review, operating-model ownership, and current pricing.
What this walkthrough proves
Technical depth connected to customer decisions.
Discovery discipline
Separate sourced facts, assumptions, missing evidence, stakeholders, and next actions.
Infrastructure fluency
Reason across model fit, throughput, latency, memory, storage, network, power, reliability, and operations.
Architecture judgment
Present a hypothesis with alternatives, risks, validation requirements, and approval boundaries.
Value engineering
Normalize operating states, preserve assumption lineage, stress the case, and distinguish modeled value from realized savings.