Track C — ML System Design

Learn to architect real, end-to-end ML systems that survive production: framing the problem, building data pipelines, training at scale, optimizing for serving, deploying safely, and monitoring for drift and fairness. Emphasis on tradeoffs, scalability, and clear communication.

%%{init: {'theme': 'base', 'themeVariables': {'primaryColor': '#F8FAFC', 'primaryBorderColor': '#F59E0B', 'lineColor': '#64748B'}}}%%
flowchart LR
  C01["C01 Framing"] --> C02["C02 Data"]
  C02 --> C03["C03 Training"]
  C03 --> C04["C04 Optimization"]
  C04 --> C05["C05 Deploy"]
  C05 --> C06["C06 Monitor"]
  C06 --> C07["C07 Capstones"]

  classDef trackC fill:#F59E0B,stroke:#0F172A,color:#0F172A
  class C01,C02,C03,C04,C05,C06,C07 trackC

Modules — Track C syllabus (7 modules)

ID	Module	What you build	Status
C01	Problem framing, metrics & decomposition	Threshold-from-costs explorer, baseline checker	DONE
C02	Data pipelines & feature engineering	Schema validator, leakage detector, split demo	DONE
C03	Training systems (fault tolerance, cost)	Checkpoint planner, throughput/cost budget, HPO budget	DONE
C04	Production model optimization	Cost model, cascade analyzer, cache simulator	DONE
C05	Deployment & serving architecture	A/B calculator, autoscaling sim, degradation planner	DONE
C06	Monitoring, drift, feedback loops & ethics	Drift detector, feedback-loop sim, fairness metrics	DONE
C07	Capstone case studies	Two full designs + their sizing calculators	DONE

Suggested order. C01 → C07 in sequence: the track is deliberately built as one arc from framing a problem to running the deployed system. If you are short on time, C01 → C04 → C07 covers the highest interview signal.

These modules have no GPU code — and that is the point

Track C ships no hip/ or cuda/ directories. The decisions here are not kernel decisions. But design material drifts easily into unfalsifiable advice, so every module is anchored by runnable, standard-library-only Python tools whose output does the teaching:

• a random split that looks fine because the leak is baked into the feature (C02)
• a cost-derived threshold that beats an F1-derived one on an unchanged model (C01)
• peeking at an A/B test 20 times taking the false-positive rate from 5% to 25% (C05)
• a recommender that raises CTR while quietly burying 60% of its catalogue (C06)

Section 5 of each module is reframed from kernel timing to system-level analysis — cost per request, latency budgets, statistical power, data quality. See CURRICULUM.md for the interview-skill mapping.