Search by concept, tool, failure, or decision.
Use / or Cmd/Ctrl + K from anywhere. Results require no server or analytics.
80 results
Engineering foundations
A churn notebook cannot be reproduced by the teammate expected to ship it. A seeded, tested, packaged, versioned training workflow with a model card.
role:platform · lifecycle:training · concept:reproducibilityCourseProblem framing, statistics & evaluation
An ETA model improves RMSE while increasing missed-delivery cost. A metric contract, leakage-safe validation, calibration analysis, and experiment plan.
lifecycle:evaluation · concept:calibration · concept:uncertaintyCourseData & feature foundations
Fraud labels arrive late and online features disagree with training data. Data contracts, point-in-time features, backfills, quality gates, and lineage.
lifecycle:data · failure:training-serving-skew · role:platformCourseLinear, probabilistic, neighbor & margin models
Support tickets need explainable routing under a 15 ms latency budget. A justified baseline across regression, Naive Bayes, kNN, and SVM.
algorithm:knn · algorithm:svm · modality:tabularCourseTrees, forests & gradient boosting
A marketplace must catch rare fraud without blocking trustworthy sellers. Calibrated random-forest and XGBoost baselines with cost-sensitive thresholds.
algorithm:random-forest · algorithm:xgboost · concept:calibrationCourseUnsupervised learning, anomalies & retrieval
Novel equipment failures appear before reliable labels exist. A PCA, clustering, anomaly, and nearest-neighbor retrieval investigation.
algorithm:pca · concept:embeddings · lifecycle:evaluationCourseForecasting & temporal ML
Regional demand shifts through promotions, holidays, and drift. Rolling validation, statistical baselines, boosted forecasts, and uncertainty bands.
modality:time-series · concept:uncertainty · failure:driftCourseRecommendation, search & ranking
Marketplace ranking learns from biased clicks and reinforces its own mistakes. Candidate generation, ranking metrics, debiasing, and online experiment design.
track:recommendation-search · failure:feedback-loop · concept:embeddingsCourseNeural networks from scratch
A nonlinear sensor classifier defeats every classical baseline. A small network built from forward pass through backpropagation and optimization.
concept:backpropagation · lifecycle:training · tool:pytorchCourseImage ML
A factory has few positive examples of costly visual defects. CNN and transfer-learning baselines with augmentation, calibration, and monitoring.
modality:image · algorithm:cnn · track:vision-multimodalCourseSequence ML: RNNs, GRUs & LSTMs
Long support conversations must be routed while preserving temporal context. A masked sequence model with BPTT, recurrent gates, and attention precursors.
modality:text · algorithm:lstm · track:nlp-llmCourseGPU-aware training
Training repeatedly OOMs and misses its nightly completion window. A profiled, mixed-precision, memory-budgeted training loop.
tool:cuda · failure:oom · lifecycle:trainingCourseDistributed & fault-tolerant training
An eight-GPU job is slower than one GPU and loses progress on failure. Efficient DDP, parallelism choices, elastic sampling, and resumable checkpoints.
concept:all-reduce · tool:pytorch · role:platformCourseBuild a transformer from scratch
Domain autocomplete needs a model whose every component the team can explain. Tokenizer, attention, RoPE, residual stack, training loop, generation, and KV cache.
concept:attention · algorithm:transformer · track:nlp-llmCourseFoundation-model lifecycle
A private support assistant must answer from evidence and resist manipulation. A grounded system design spanning adaptation, RAG, evaluation, and safety.
concept:rag · concept:evaluation · failure:prompt-injectionCourseFrontier architectures & efficient inference
Long-context generation breaks latency and cost targets. A capacity plan using efficient attention, quantization, batching, and modern routing.
concept:kv-cache · tool:vllm · lifecycle:inferenceCourseProduction data & ML platforms
A late feature silently corrupts every downstream prediction. A traced batch/stream platform with orchestration, registry, lineage, and ownership.
tool:kafka · tool:mlflow · role:platformCourseServing & inference systems
Launch traffic causes p99 spikes and leaves expensive GPUs idle. Batch and online paths with batching, autoscaling, load tests, canaries, and rollback.
lifecycle:inference · tool:triton · failure:latencyCourseReliability, monitoring & experimentation
Model quality decays while every infrastructure dashboard stays green. Joined data/model/service telemetry, drift policy, experiments, alerts, and runbooks.
failure:drift · lifecycle:monitoring · role:reliabilityCourseGovernance, fairness, privacy & security
A high-performing system creates subgroup harm and leaks sensitive information. A launch review covering fairness, privacy, licensing, abuse, security, and response.
concept:fairness · concept:privacy · failure:supply-chainLessonFrom notebook to reproducible run
A reproducible result is a result whose complete input identity can be replayed.
role:platform · lifecycle:training · concept:reproducibilityLessonContracts before models
Contracts keep bad meaning from becoming good-looking features.
role:platform · lifecycle:training · concept:reproducibilityLessonTraining code is product code
The shipped model is the reloaded artifact, not the object still living in the trainer.
role:platform · lifecycle:training · concept:reproducibilityLessonContainers and configuration without surprises
A container is reproducible only when its image, inputs, configuration, and outputs all have explicit identities.
role:platform · lifecycle:training · concept:reproducibilityLessonObservable, idempotent, failure-safe workflows
A safe workflow makes commitment atomic and repetition uneventful.
role:platform · lifecycle:training · concept:reproducibilityLessonRead a research paper as an executable specification
Turn a research paper into a testable claim map, implementation contract, and prioritized reading plan before writing code.
role:platform · lifecycle:training · concept:reproducibilityLessonReproduce, validate, and productionize a paper
Implement a paper from baseline to novel component, diagnose reproduction gaps, and adapt it for production without silently changing its claim.
role:platform · lifecycle:training · concept:reproducibilityLessonPredict decisions, not just labels
A model score earns its place by improving a declared decision.
lifecycle:evaluation · concept:calibration · concept:uncertaintyLessonLabels, availability, and leakage
Train on what the system knew then, label with what became known later.
lifecycle:evaluation · concept:calibration · concept:uncertaintyLessonBaselines and honest splits
A baseline sets the floor; an honest split keeps the ruler straight.
lifecycle:evaluation · concept:calibration · concept:uncertaintyLessonMetrics, thresholds, and calibration
Measure ordering, repair probability meaning, then choose the policy the operation can actually run.
lifecycle:evaluation · concept:calibration · concept:uncertaintyLessonUncertainty and online experiments
Uncertainty tempers the estimate; randomization tests the cause; guardrails protect the product.
lifecycle:evaluation · concept:calibration · concept:uncertaintyLessonMetric selection by task and decision
Select metrics from the prediction target, decision, error costs, prevalence, slices, horizon, and deployment constraints—not from habit.
lifecycle:evaluation · concept:calibration · concept:uncertaintyLessonData contracts, ownership, and lineage
A trustworthy dataset explains meaning, accountability, and ancestry—not just columns.
lifecycle:data · failure:training-serving-skew · role:platformLessonPoint-in-time-correct features
Point-in-time correctness preserves the knowledge boundary of the original decision.
lifecycle:data · failure:training-serving-skew · role:platformLessonFeature transformations that survive production
Ship the fitted representation contract with the model, in the same order and with the same edge cases.
lifecycle:data · failure:training-serving-skew · role:platformLessonBatch, stream, and training-serving parity
Parity starts with one temporal meaning, then proves every execution path implements it.
lifecycle:data · failure:training-serving-skew · role:platformLessonLabels, snapshots, and backfills
A label snapshot records what was knowable by a date, not the database’s latest opinion.
lifecycle:data · failure:training-serving-skew · role:platformLessonLinear regression as a production baseline
A linear baseline is valuable because its assumptions, errors, and serving cost are easy to see.
algorithm:knn · algorithm:svm · modality:tabularLessonLogistic regression, odds, and calibration
Logistic regression adds evidence in log-odds space; calibration and policy give that score operational meaning.
algorithm:knn · algorithm:svm · modality:tabularLessonNaive Bayes under sparse data
Naive Bayes turns sparse counts into additive evidence—quickly, smoothly, and with assumptions in plain sight.
algorithm:knn · algorithm:svm · modality:tabularLessonkNN, distance, and local decisions
In kNN, representation defines the neighborhood, and the neighborhood defines the decision.
algorithm:knn · algorithm:svm · modality:tabularLessonSVMs, margins, and kernels
SVMs learn from the boundary’s hardest points; kernels change geometry, not scores into probabilities.
algorithm:knn · algorithm:svm · modality:tabularLessonTrees as learned decision rules
A tree learns readable local rules; constraints keep those rules from becoming a lookup table.
algorithm:random-forest · algorithm:xgboost · concept:calibrationLessonRandom forests and diversity
A forest is strong when its trees are individually useful and collectively disagree in the right ways.
algorithm:random-forest · algorithm:xgboost · concept:calibrationLessonGradient boosting as error correction
Boosting learns by repeatedly asking: what structured error remains?
algorithm:random-forest · algorithm:xgboost · concept:calibrationLessonXGBoost: regularized boosting at scale
XGBoost scales boosting by turning gradient statistics into regularized decisions about leaves and splits.
algorithm:random-forest · algorithm:xgboost · concept:calibrationLessonCalibrating and explaining ensembles
Calibrate probabilities for decisions, and read explanations as model accounting—not causal proof.
algorithm:random-forest · algorithm:xgboost · concept:calibrationLessonPCA and representation geometry
PCA preserves dominant geometric variation, not automatically the information your decision needs.
algorithm:pca · concept:embeddings · lifecycle:evaluationLessonClustering as an operational hypothesis
A cluster becomes knowledge only when stable geometry connects to a repeatable human or system action.
algorithm:pca · concept:embeddings · lifecycle:evaluationLessonDensity clusters and novel anomalies
Density methods flag what is locally disconnected; people and outcomes determine what is truly novel and important.
algorithm:pca · concept:embeddings · lifecycle:evaluationLessonEmbeddings and similarity
Similarity is a modeling choice: the same vectors tell different stories under cosine and dot product.
algorithm:pca · concept:embeddings · lifecycle:evaluationLessonApproximate retrieval under latency
Approximate retrieval is an explicit recall–latency contract, not a transparent substitute for exact search.
algorithm:pca · concept:embeddings · lifecycle:evaluationLessonForecast contracts and naive models
A forecast begins with two clocks—when you know and when you predict—and a baseline that respects both.
modality:time-series · concept:uncertainty · failure:driftLessonTrend, seasonality, and known-future signals
A temporal feature is valid only if its value could be reconstructed exactly at the forecast origin.
modality:time-series · concept:uncertainty · failure:driftLessonBacktesting without time travel
A faithful backtest is a replay of past information states, not merely a sorted split.
modality:time-series · concept:uncertainty · failure:driftLessonStatistical and ML forecasters
Choose a forecaster and a multi-step strategy together; the estimator alone does not define the forecast.
modality:time-series · concept:uncertainty · failure:driftLessonIntervals, drift, and forecast operations
Forecast reliability is the joint discipline of calibrated ranges, drift evidence, and a usable fallback.
modality:time-series · concept:uncertainty · failure:driftLessonFrom corpus to slate: retrieve, score, re-rank
Diagnose recommendation quality stage by stage: coverage first, ordering second, slate policy third.
track:recommendation-search · failure:feedback-loop · concept:embeddingsLessonClicks are not labels
A click is an exposed, policy-shaped outcome—not a universal relevance label.
track:recommendation-search · failure:feedback-loop · concept:embeddingsLessonBM25, factorization, two towers, and ANN
Hybrid retrieval works because lexical, collaborative, and content representations fail in different places.
track:recommendation-search · failure:feedback-loop · concept:embeddingsLessonPointwise, pairwise, and listwise ranking
Pointwise learns values, pairwise learns preferences, and listwise learns positions—choose the supervision that matches the slate.
track:recommendation-search · failure:feedback-loop · concept:embeddingsLessonServing freshness, diversity, and experiments
A recommendation ships only when relevance, freshness, diversity, and system guardrails succeed together.
track:recommendation-search · failure:feedback-loop · concept:embeddingsLessonBeyond linear boundaries
Hidden nonlinearities earn their place by making the final decision simpler in a learned representation.
concept:backpropagation · lifecycle:training · tool:pytorchLessonForward pass, activations, losses, and initialization
A trustworthy forward pass makes shapes, activation semantics, initialization, and loss expectations explicit.
concept:backpropagation · lifecycle:training · tool:pytorchLessonBackpropagation without magic
Backprop becomes ordinary once every edge in the forward graph has a local derivative you can inspect.
concept:backpropagation · lifecycle:training · tool:pytorchLessonSGD, momentum, AdamW, and schedules
The trained state is parameters plus optimizer history, schedule position, data order, and randomness.
concept:backpropagation · lifecycle:training · tool:pytorchLessonRegularize, normalize, debug, and ship
Shipping a network means freezing its inference behavior, not merely saving its weights.
concept:backpropagation · lifecycle:training · tool:pytorchLessonActivation functions: geometry, gradients, and failure modes
Select activation functions by signal range, gradient flow, architecture, precision, and failure mode; then diagnose saturation and dead units from telemetry.
concept:backpropagation · lifecycle:training · tool:pytorchLessonLoss functions by task, data, and deployment constraints
Choose and debug training objectives as optimization surrogates for the real task, including imbalance, outliers, uncertainty, ranking, dense prediction, and generation.
concept:backpropagation · lifecycle:training · tool:pytorchLessonImages are data contracts
Image preprocessing is model code; version and test it from raw bytes to final tensor.
modality:image · algorithm:cnn · track:vision-multimodalLessonConvolution under the microscope
Convolution is a repeated local dot product whose shape and spatial behavior follow directly from kernel, padding, stride, and dilation.
modality:image · algorithm:cnn · track:vision-multimodalLessonReceptive fields, normalization, and residuals
Depth works when context grows deliberately and identity paths keep shapes and gradients connected.
modality:image · algorithm:cnn · track:vision-multimodalLessonTransfer learning for classification, detection, and segmentation
Transfer the representation in measured stages, while letting the required output and leakage-safe split define the task.
modality:image · algorithm:cnn · track:vision-multimodalLessonVision in the wild: thresholds, latency, and drift
Production vision is a thresholded, timed, monitored decision system—not an accuracy number attached to a model file.
modality:image · algorithm:cnn · track:vision-multimodalLessonSequences are not IID
In sequence learning, chronology is part of the feature schema: define the cutoff first, then construct inputs, labels, and splits around it.
modality:text · algorithm:lstm · track:nlp-llmLessonRecurrent state and BPTT
Recurrent state carries information forward; BPTT carries responsibility backward, and truncation decides how far that responsibility can travel.
modality:text · algorithm:lstm · track:nlp-llmLessonGates that remember: LSTM and GRU
Gates do not create memory by magic; they create learnable paths for preserving, replacing, and exposing state.
modality:text · algorithm:lstm · track:nlp-llmLessonMany-to-one and many-to-many training
Choose the output pattern from the target, then make padding mathematically invisible to both state selection and loss.
modality:text · algorithm:lstm · track:nlp-llmLessonStateful streaming inference without future leakage
For streaming inference, the state transition, ordering policy, deduplication rule, and model version are all part of the model.
modality:text · algorithm:lstm · track:nlp-llm