Day 45: Stage 1 capstone: ≥90% val accuracy, micrograd defense
Stage 1 capstone: the exit criteria
Stage 1 is done when its exit criteria pass — not when the code ran once. Three checks, straight from the roadmap: you can defend backprop from scratch, the classifier hits ≥90% validation accuracy on ≥8 classes with overfitting handled and curves shown, and the model is exported and served behind NestJS with a recorded baseline latency.
Exit criterion 1: whiteboard backprop
Defend the fundamentals, unprompted
Without notes: implement micrograd-style backprop for a scalar expression, and whiteboard the chain rule through a 2-layer net (Day 25's narrative). If you can do this cold, you've cleared the conceptual bar that most 'used a model once' candidates cannot — it's the single most differentiating thing you learned this stage.
Exit criterion 2: ≥90% val accuracy, defensibly
The model checks
Exit criterion 3: served and measured
The serving checks
Day-90 is closer than it looks
Stages 0–2 plus your existing backend/K8s résumé make you interviewable for CV-flavored AI Engineer roles at Day 90. The classifier you just shipped — trained, measured, served, defensible — is already a portfolio piece most candidates don't have. Write this week's log (built / measured / broke / verified-vs-assumed); it's interview material, not busywork.
Your Garment Classifier reaches 91% validation accuracy but you achieved it by tuning until the validation number looked good, reusing the validation set many times. What is the risk, per the roadmap's ethos?
