Projects for XP

A retrieval pipeline that chunks documents, searches by keyword and meaning, reranks results and measures its own quality.

1. 1. Split documents into overlapping chunksChunking Documents for RAG
2. 2. Turn token vectors into one embedding per chunkSentence Embeddings: Masked Mean Pooling
3. 3. Find the closest chunks to a queryEmbedding Retrieval with Cosine Similarity
4. 4. Add keyword search for exact termsBM25 Keyword Search
5. 5. Rerank the shortlist for precisionTwo-Stage Retrieval with a Reranker
6. 6. Score the retriever with recall and MRRRAG Retrieval Metrics: Recall@k, MRR, nDCG

The sampling loop at the heart of every chatbot: temperature, nucleus and min-p sampling, banned tokens and best-of-n.

1. 1. Turn logits into probabilitiesSoftmax with Temperature
2. 2. Block tokens you never wantLogit Bias and Banned Tokens
3. 3. Sample from the nucleusTop-k & Top-p (Nucleus) Filtering
4. 4. Add min-p as a safer cut-offMin-p Sampling
5. 5. Chain it all into one pipelineThe Full Sampling Pipeline
6. 6. Pick the best of several samplesBest-of-N Sampling with a Reward Model

A transformer attention layer from positional encodings to multi-head attention, and the gradients to train it.

1. 1. Tell the model where each token isSinusoidal Positional Encoding
2. 2. A softmax that never overflowsNumerically Stable Softmax
3. 3. Stop tokens peeking at the futureBuilding Attention Masks
4. 4. Multi-head attention, forward passMulti-Head Self-Attention Forward Pass
5. 5. Backpropagate through attentionBackpropagate Through Attention

A text classifier with features, a model, honest validation and a tuned decision threshold.

1. 1. Turn emails into TF-IDF featuresTF-IDF Vectoriser
2. 2. Train a Naive Bayes modelNaive Bayes Text Classifier
3. 3. Validate it with k-fold CVK-Fold Cross-Validation Splits
4. 4. Measure precision, recall and F1Precision, Recall & F1
5. 5. Compare models with ROC AUCROC AUC from Scratch
6. 6. Pick a threshold that fits the costsChoose a Threshold by Business Cost

A complete experiment analysis: clean data, the right test, effect size, confidence and the traps to avoid.

1. 1. Compute conversion through the funnelFunnel Conversion from Event Logs
2. 2. Find outliers before they skew resultsOutlier Detection with the IQR Rule
3. 3. Test the difference in meansWelch's t-Test for Two Means
4. 4. Report how big the effect isEffect Size: Cohen's d
5. 5. Put a confidence interval on conversionConfidence Interval for a Conversion Rate (Wilson)
6. 6. Show why peeking inflates false positivesThe Peeking Problem in A/B Tests

The queries behind a sales dashboard: revenue, retention, cohorts and churn.

1. 1. Revenue by countryRevenue by Country
2. 2. Monthly active customersMonthly Active Customers
3. 3. A running revenue totalRunning Total per Customer
4. 4. Repeat purchase rateRepeat Purchase Rate
5. 5. Orders by signup cohortOrders per Customer by Signup Cohort
6. 6. Who has churnedFind Lapsed Customers

The plumbing of a production client: query strings, retries, caching, rate limiting and readable logs.

1. 1. Parse and build query stringsParse a URL Query String
2. 2. Retry failures with backoffRetry Decorator with Exponential Backoff
3. 3. Cache responses for a whileMemoize with a Time-to-Live
4. 4. Stay under the rate limitSliding-Window Rate Limiter per User
5. 5. Parse the logs it writesParse and Summarise Log Files

Grid and graph pathfinding, from flood fill to A*, plus scheduling the trip.

1. 1. Find connected regions on a mapNumber of Islands
2. 2. Shortest path on a gridShortest Path in a Grid
3. 3. Order stops with dependenciesTask Ordering with Topological Sort
4. 4. Distances between every pair of placesAll-Pairs Shortest Paths (Floyd–Warshall)
5. 5. Fast routes with A*A* Pathfinding on a Grid