Showcase Architecture

This note maps related showcases into cohesive, in-repo learning tracks.

Why this architecture

• Keep each showcase focused on one learning outcome.
• Preserve reproducibility and short demo runtime.
• Avoid monolithic project structure for students.

Ranking Track

1. projects/learning-to-rank-foundations-showcase
2. Grouped ranking data preparation and relevance labeling.
3. LambdaRank model training.

4. NDCG-focused evaluation and split artifacts.

5. projects/ranking-api-productization-showcase

6. FastAPI ranking endpoints (/health, /model/schema, /score, /rank).
7. Model artifact loading and schema-safe scoring.
8. Structured request logging and OpenAPI export workflow.

Forecasting And Observability Track

1. projects/nyc-demand-forecasting-foundations-showcase
2. TLC-style hourly aggregation and time feature engineering.
3. Explicit time-ordered train/val/test split.
4. Demand forecasting metrics (MAE, RMSE, sMAPE).

5. Optional real TLC download path with synthetic default mode.

6. projects/demand-api-observability-showcase

7. FastAPI demand serving endpoint (/predict) and health checks.
8. Prometheus metrics endpoint (/metrics) and request latency counters.
9. Optional OpenTelemetry instrumentation hooks.
10. OpenAPI export/check and API behavior tests.

Agentic Experimentation Track

1. projects/autoresearch
2. Fixed-budget autonomous research loop for tiny language-model pretraining.
3. Unified platform guidance for autoresearch-macos and autoresearch.
4. Decision artifacts for keep/discard/crash outcomes.

5. Codex and Claude Code launch briefs tied to the upstream repos.

6. projects/agentic-course-assistant-showcase

7. Deterministic course-assistant workflow for routing, tools, guardrails, and trace inspection.
8. Optional OpenAI Agents SDK and Google ADK reference modules.
9. Stable local artifacts that make agent decisions inspectable before live API usage.

Sequential Decision Track

1. projects/rl-bandits-policy-showcase
2. Small exploration-versus-exploitation warm-up with reward and regret traces.

3. Stationary multi-armed bandit policies that stay easy to benchmark.

4. projects/student-support-rl-showcase

5. Small synthetic student-support MDP with readable transitions.
6. Real contextual bandit warm-up plus random, heuristic, tabular Q-learning, and optional DQN/PPO comparison.
7. Reward-hacking audit, governance notes, and deploy/shadow/reject memo.

Intentional Scope Boundaries

• Full-size raw datasets are excluded to keep clone and run workflows lightweight.
• Large generated caches are excluded from version control.
• Each showcase keeps only teaching-critical components and artifacts.