AI Venture Build · Capstone brief
VEN-03 · Does this need a model?
Design the version of your product that uses no machine learning at all — a form, a spreadsheet, a rules engine, an off-the-shelf tool, a person doing the work. Then say what the model has to beat, on a named metric, to earn its place. If it does not beat the simple version, that is your answer.
The question
Design the version of your product that uses no machine learning at all — a form, a spreadsheet, a rules engine, an off-the-shelf tool, a person doing the work. Then say what the model has to beat, on a named metric, to earn its place. If it does not beat the simple version, that is your answer.
System / materials
The venture concept and its core job. A genuinely serious non-ML design — the teacher should reject a straw man, because a deliberately bad manual version makes the whole comparison worthless. A named decision metric agreed in advance (accuracy on a defined task, time saved per unit, cost per unit, or coverage), plus a rough cost for both paths. Where feasible, run the manual version on a handful of real cases; concierge-style manual delivery is often the honest MVP and is a legitimate outcome.
Expected failure modes
Building a straw-man manual version so ML wins. Choosing a metric after seeing which way it went. Ignoring what the manual version buys you — control, explainability, no inference cost, no model drift, no vendor dependency. Assuming the model gets better and the manual version does not. Adding ML because the competition category or the judges are expected to reward it.
Done looks like
A comparison memo: the non-ML design described well enough that someone could actually run it; the pre-named metric with the bar the model must clear; both paths costed at least roughly, including the human time the manual version needs; and a decision with reasoning — including "not yet" or "not at all," which are full-credit outcomes in this bank.
Five C's
CT: separating "a model could do this" from "a model should do this." CR: designing an honest simple version. CO: a peer argues the manual path and is answered on the metric. CM: a decision memo a judge cannot dismiss as hand-waving. CZ: who bears the cost of an unnecessary model — the customer in price, or the user in unexplainable errors.
Mentor role
A founder, engineer, or operations lead reviews the non-ML design before the comparison. Standing instruction: reject a straw-man manual version. School-supervised.
Rubric calibration
R1: one job, one metric, two designs. R2: the manual design is reproducible. R3: this brief is R3 — the baseline is the deliverable. R4: model-specific risks named. R5: memo is decision-ready. R6: willing to conclude the model is not needed.
Two ways this goes wrong
(a) A manual version nobody would ever run, so the comparison is theater. (b) The metric is chosen after the result, so the conclusion was never at risk.
Credit lane fit
Lane A immediately. The single most useful brief in this bank for a student whose idea is "AI for X." No verified credit claim.