Build the pipeline. Ship the system.

Build a versioned, validated document dataset from raw rows.

1. L1Row Contract & Validator~12 min›
2. L2Deduplication + Provenance~15 min›
3. L3Group-Safe Split~18 min›
4. L4Incremental Ingest + Dead Letters~20 min›
5. L5Rebuildable Dataset Release~25 min›

Build metrics and a baseline that earns the right to be improved.

1. L1Confusion Matrix~10 min›
2. L2Precision, Recall, F1~12 min›
3. L3Majority + Keyword Baselines~15 min›
4. L4Slices + Thresholds~18 min›
5. L5Evaluation Gate~22 min›

Build logistic classification without a model library.

1. L1Score Function~12 min›
2. L2Thresholded Prediction~14 min›
3. L3Sigmoid + Log Loss~16 min›
4. L4Gradient Descent~22 min›
5. L5Trainable Artifact~28 min›

Build a reproducible runner and error-analysis workflow.

1. L1Reproducible Run Config~12 min›
2. L2Experiment Runner~18 min›
3. L3Ablation + Comparison~18 min›
4. L4Error Taxonomy~20 min›
5. L5Evidence-Based Selection~24 min›

Build restartable ingestion, cleaning, lineage, and drift detection.

1. L1Source Adapter~12 min›
2. L2Checkpointed Pipeline~16 min›
3. L3Retry + Dead Letter~18 min›
4. L4Rebuild from Manifest~20 min›
5. L5Drift Warning~24 min›

Build a typed, bounded inference service around the classifier.

1. L1Request Contract~12 min›
2. L2Model Adapter~15 min›
3. L3Request Dispatcher~18 min›
4. L4Health + Readiness~16 min›
5. L5Batch + Lifecycle Trace~22 min›

Build a queue-backed worker with safe retries and recovery.

1. L1Queue State~12 min›
2. L2Worker Loop~15 min›
3. L3Backoff + Dead Letter~18 min›
4. L4Idempotency~18 min›
5. L5Checkpoint + Backpressure~24 min›

Build an inverted-index search engine and evaluation set.

1. L1Tokenizer~10 min›
2. L2Inverted Index~16 min›
3. L3BM25-Style Score~18 min›
4. L4Query + Freshness~20 min›
5. L5Retrieval Evaluation~22 min›

Build deterministic embeddings and nearest-neighbor search without a vector database.

1. L1Deterministic Encoder~14 min›
2. L2Cosine Similarity~12 min›
3. L3Exact Top-k~16 min›
4. L4Versioned Updates~20 min›
5. L5Semantic Evaluation~24 min›

Separate candidate generation, feature scoring, and reranking.

1. L1Candidate Merge~12 min›
2. L2Ranking Features~16 min›
3. L3Reranker~18 min›
4. L4Diversity + Freshness~18 min›
5. L5Ranking Evaluation~24 min›

Build citation-backed answers with abstention.

1. L1Context Selector~14 min›
2. L2Citation Binder~18 min›
3. L3Structured Answer Validator~16 min›
4. L4Abstention + Safety~20 min›
5. L5End-to-End Service~28 min›

Build a tiny tensor and automatic-differentiation runtime.

1. L1Tensor Shape Contract~16 min›
2. L2Matmul + Broadcasting~22 min›
3. L3Computation Graph~20 min›
4. L4Backward Pass~24 min›
5. L5Gradient Check + Optimizer~28 min›

Build a tiny causal language model with inspectable attention.

1. L1Tokenizer + Batches~16 min›
2. L2Embeddings + Positions~20 min›
3. L3Masked Self-Attention~26 min›
4. L4Block + Train Step~30 min›
5. L5Sampling + Failure Report~32 min›

Build registry, promotion, rollout, observability, and rollback.

1. L1Model Registry~16 min›
2. L2Promotion Gate~20 min›
3. L3Rollout + Rollback~22 min›
4. L4Observability + SLO~24 min›
5. L5Incident Recovery~28 min›

Integrate artifacts from Projects 1–14 into one reviewable system.

1. L1Ingest to Index~24 min›
2. L2Train to Registry~26 min›
3. L3Unified Service~28 min›
4. L4Failure + Recovery~30 min›
5. L5Design Review + Release~35 min›