MLE/ROADMAP
/Search
Menu
Connected field reference

Look up the invariant. Follow it into production.

Every entry routes back to the lessons, assignments, and cases where it changes a real system decision.

20 formal anchors

Formulas with declared semantics

Engineering foundations

Content-addressed run identity

run_id⁡=SHA256⁡(code∥data∥image∥config∥seed)\operatorname{run\_id} = \operatorname{SHA256}(\mathrm{code} \mathbin{\|} \mathrm{data} \mathbin{\|} \mathrm{image} \mathbin{\|} \mathrm{config} \mathbin{\|} \mathrm{seed})
code
source revision digest
data
immutable dataset digest
image
OCI image digest
seed
declared random seed
Open formal lesson ↗
Evaluation

Cost-aware threshold selection

C(t)=cFN FN(t)+cFP FP(t)C(t) = c_{\mathrm{FN}}\,\mathrm{FN}(t) + c_{\mathrm{FP}}\,\mathrm{FP}(t)
t
decision threshold
FN, FP
error counts at t
c_FN, c_FP
business cost per error
Open formal lesson ↗
Data & features

Feature availability invariant

tavailable(fi)≤tdecisionfor every feature fit_{\mathrm{available}}(f_i) \le t_{\mathrm{decision}} \quad \text{for every feature } f_i
t_available
time the value became knowable
t_decision
prediction cutoff
f_i
feature i
Open formal lesson ↗
Classical baselines

Logistic probability

p(y=1∣x)=11+exp⁡(−x⊤β)p(y=1 \mid x) = \frac{1}{1 + \exp(-x^\top \beta)}
x
feature vector
β
learned coefficients
xᵀβ
linear log-odds score
Open formal lesson ↗
Trees & boosting

Stage-wise boosting update

Fm(x)=Fm−1(x)+ηhm(x)F_m(x) = F_{m-1}(x) + \eta h_m(x)
F_m
ensemble after stage m
h_m
new weak learner
η
learning rate
Open formal lesson ↗
Unsupervised & retrieval

Cosine similarity

cos⁡(u,v)=u⊤v∥u∥2∥v∥2\cos(u,v) = \frac{u^\top v}{\lVert u \rVert_2 \lVert v \rVert_2}
u, v
embedding vectors
uᵀv
dot product
∥·∥₂
vector magnitude
Open formal lesson ↗
Forecasting

Quantile forecast objective

q^τ=arg min⁡q∑max⁡ ⁣(τ(y−q),(τ−1)(y−q))\hat{q}_\tau = \operatorname*{arg\,min}_q \sum \max\!\left(\tau(y-q),(\tau-1)(y-q)\right)
τ
target quantile
q̂_τ
predicted quantile
y
observed value
Open formal lesson ↗
Ranking systems

Discounted cumulative gain

DCG@K=∑r=1K2relr−1log⁡2(r+1)\mathrm{DCG}@K = \sum_{r=1}^{K} \frac{2^{\mathrm{rel}_r}-1}{\log_2(r+1)}
rel_r
relevance at rank r
r
one-indexed rank
K
slate cutoff
Open formal lesson ↗
Neural networks

Backpropagated error

δl=(Wl+1⊤δl+1)⊙φ′(zl)\delta_l = \left(W_{l+1}^\top \delta_{l+1}\right) \odot \varphi'(z_l)
δ_l
loss gradient at layer l
W
weight matrix
φ′
activation derivative
Open formal lesson ↗
Vision

Convolution output size

out=⌊in+2p−d(k−1)−1s+1⌋\mathrm{out} = \left\lfloor \frac{\mathrm{in} + 2p - d(k-1) - 1}{s} + 1 \right\rfloor
in, out
spatial input/output size
k
kernel size
s
stride
p
padding
d
dilation
Open formal lesson ↗
Sequence models

LSTM cell-state update

ct=ft⊙ct−1+it⊙gtc_t = f_t \odot c_{t-1} + i_t \odot g_t
c_t
cell state
f_t
forget gate
i_t
input gate
g_t
candidate state
Open formal lesson ↗
GPU training

Roofline bound

attainable FLOP/s≤min⁡ ⁣(Ppeak,I⋅Bmem)\text{attainable FLOP/s} \le \min\!\left(P_{\mathrm{peak}}, I \cdot B_{\mathrm{mem}}\right)
P_peak
peak compute
I
arithmetic intensity
B_mem
memory bandwidth
Open formal lesson ↗
Distributed training

Scaling efficiency

E=throughputWW⋅throughput1E = \frac{\mathrm{throughput}_W}{W \cdot \mathrm{throughput}_1}
W
worker/GPU count
throughput_W
distributed samples per second
throughput_1
single-worker baseline
Open formal lesson ↗
Transformer from scratch

Scaled dot-product attention

Attention⁡(Q,K,V)=softmax⁡ ⁣(QK⊤dk+M)V\operatorname{Attention}(Q,K,V) = \operatorname{softmax}\!\left(\frac{QK^\top}{\sqrt{d_k}} + M\right)V
Q, K, V
query, key, value matrices
d_k
head width
M
causal or padding mask
Open formal lesson ↗
Foundation models

Autoregressive next-token loss

LNLL(θ)=−1T∑tlog⁡pθ ⁣(xt∣x<t)\mathcal{L}_{\mathrm{NLL}}(\theta) = -\frac{1}{T}\sum_t \log p_\theta\!\left(x_t \mid x_{<t}\right)
T
sequence length
x_t
target token
θ
model parameters
Open formal lesson ↗
Efficient inference

KV-cache memory estimate

MKV≈2⋅L⋅B⋅S⋅Hkv⋅dh⋅qM_{\mathrm{KV}} \approx 2 \cdot L \cdot B \cdot S \cdot H_{\mathrm{kv}} \cdot d_h \cdot q
L
decoder layers
B
active sequences
S
cached tokens
H_kv
KV heads
q
bytes per scalar
Open formal lesson ↗
ML platforms

Point-in-time feature lookup

fj(tp)=value⁡ ⁣(arg max⁡etevent(e))subject to tevent(e), tavailable(e)≤tpf_j(t_p) = \operatorname{value}\!\left(\operatorname*{arg\,max}_e t_{\mathrm{event}}(e)\right) \quad \text{subject to } t_{\mathrm{event}}(e),\,t_{\mathrm{available}}(e) \le t_p
j
entity
t_p
historical prediction time
e
eligible event
Open formal lesson ↗
Serving systems

Queue utilization and Little’s law

ρ=λcμandN=λW\rho = \frac{\lambda}{c\mu} \quad \text{and} \quad N = \lambda W
λ
arrival rate
c
worker count
μ
service rate per worker
W
mean time in system
Open formal lesson ↗
Reliability

Difference-in-means experiment estimate

τ^=YˉT−YˉC,SE=sT2nT+sC2nC\hat{\tau} = \bar{Y}_T - \bar{Y}_C,\quad \mathrm{SE} = \sqrt{\frac{s_T^2}{n_T} + \frac{s_C^2}{n_C}}
Ȳ_T, Ȳ_C
treatment/control means
s²
sample variance
n
independent randomized units
Open formal lesson ↗
Responsible ML

Differential privacy guarantee

Pr⁡[M(D)∈S]≤exp⁡(ε)Pr⁡[M(D′)∈S]+δ\Pr[M(D) \in S] \le \exp(\varepsilon)\Pr[M(D') \in S] + \delta
M
randomized mechanism
D, D′
adjacent datasets
ε, δ
privacy-loss parameters
Open formal lesson ↗
Model families

Algorithms are conditional choices.

algorithm:cnn

CNN

CNN in the context of production machine-learning systems.

algorithm:knn

kNN

kNN in the context of production machine-learning systems.

algorithm:lstm

LSTM

LSTM in the context of production machine-learning systems.

algorithm:pca

PCA

PCA in the context of production machine-learning systems.

algorithm:random-forest

Random Forest

Random Forest in the context of production machine-learning systems.

algorithm:svm

SVM

SVM in the context of production machine-learning systems.

algorithm:transformer

Transformer

Transformer in the context of production machine-learning systems.

algorithm:xgboost

XGBoost

Regularized, second-order gradient-boosted decision trees designed for efficient tabular learning.

Concrete systems

Tools, with the boundary they own.

tool:cuda

CUDA

CUDA in the context of production machine-learning systems.

tool:kafka

Kafka

Kafka in the context of production machine-learning systems.

tool:mlflow

MLflow

MLflow in the context of production machine-learning systems.

tool:pytorch

PyTorch

A tensor, autograd, neural-network, and distributed-training runtime.

tool:triton

Triton

Triton in the context of production machine-learning systems.

tool:vllm

vLLM

vLLM in the context of production machine-learning systems.

Controlled vocabulary

Glossary by type:name.

CNNalgorithm:cnn
CNN in the context of production machine-learning systems.
kNNalgorithm:knn
kNN in the context of production machine-learning systems.
LSTMalgorithm:lstm
LSTM in the context of production machine-learning systems.
PCAalgorithm:pca
PCA in the context of production machine-learning systems.
Random Forestalgorithm:random-forest
Random Forest in the context of production machine-learning systems.
SVMalgorithm:svm
SVM in the context of production machine-learning systems.
Transformeralgorithm:transformer
Transformer in the context of production machine-learning systems.
XGBoostalgorithm:xgboost
Regularized, second-order gradient-boosted decision trees designed for efficient tabular learning.
Ablation Logicconcept:ablation-logic
Ablation Logic in the context of production machine-learning systems.
Acceptance Gatesconcept:acceptance-gates
Acceptance Gates in the context of production machine-learning systems.
Activation Memoryconcept:activation-memory
Activation Memory in the context of production machine-learning systems.
Adamwconcept:adamw
Adamw in the context of production machine-learning systems.
Affine Layersconcept:affine-layers
Affine Layers in the context of production machine-learning systems.
Airflow Dagsterconcept:airflow-dagster
Airflow Dagster in the context of production machine-learning systems.
All Reduceconcept:all-reduce
All Reduce in the context of production machine-learning systems.
Anomaly Triageconcept:anomaly-triage
Anomaly Triage in the context of production machine-learning systems.
Arithmetic Intensityconcept:arithmetic-intensity
Arithmetic Intensity in the context of production machine-learning systems.
As Of Joinsconcept:as-of-joins
As Of Joins in the context of production machine-learning systems.
Assumption Ledgersconcept:assumption-ledgers
Assumption Ledgers in the context of production machine-learning systems.
Attentionconcept:attention
Content-addressed mixing of token representations through query, key, and value projections.
Augmentationconcept:augmentation
Augmentation in the context of production machine-learning systems.
Autoscalingconcept:autoscaling
Autoscaling in the context of production machine-learning systems.
Availability Timeconcept:availability-time
Availability Time in the context of production machine-learning systems.
Backpropagationconcept:backpropagation
Backpropagation in the context of production machine-learning systems.
Baggingconcept:bagging
Bagging in the context of production machine-learning systems.
Baseline Contractsconcept:baseline-contracts
Baseline Contracts in the context of production machine-learning systems.
Batchingconcept:batching
Batching in the context of production machine-learning systems.
Bayes Ruleconcept:bayes-rule
Bayes Rule in the context of production machine-learning systems.
Benchmark Validityconcept:benchmark-validity
Benchmark Validity in the context of production machine-learning systems.
Bprconcept:bpr
Bpr in the context of production machine-learning systems.
Bpttconcept:bptt
Bptt in the context of production machine-learning systems.
Build Vs Buyconcept:build-vs-buy
Build Vs Buy in the context of production machine-learning systems.
Calibrationconcept:calibration
Calibration in the context of production machine-learning systems.
Camera Driftconcept:camera-drift
Camera Drift in the context of production machine-learning systems.
Canariesconcept:canaries
Canaries in the context of production machine-learning systems.
Canaryconcept:canary
Canary in the context of production machine-learning systems.
Candidate Generationconcept:candidate-generation
Candidate Generation in the context of production machine-learning systems.
Causal Masksconcept:causal-masks
Causal Masks in the context of production machine-learning systems.
Censoringconcept:censoring
Censoring in the context of production machine-learning systems.
Chain Ruleconcept:chain-rule
Chain Rule in the context of production machine-learning systems.
Champion Challengerconcept:champion-challenger
Champion Challenger in the context of production machine-learning systems.
Checkpoint Reshardingconcept:checkpoint-resharding
Checkpoint Resharding in the context of production machine-learning systems.
Checkpointingconcept:checkpointing
Checkpointing in the context of production machine-learning systems.
Circuit Breakersconcept:circuit-breakers
Circuit Breakers in the context of production machine-learning systems.
Claim Evidence Mapsconcept:claim-evidence-maps
Claim Evidence Maps in the context of production machine-learning systems.
Class Weightingconcept:class-weighting
Class Weighting in the context of production machine-learning systems.
Cold Startconcept:cold-start
Cold Start in the context of production machine-learning systems.
Compilationconcept:compilation
Compilation in the context of production machine-learning systems.
Confidence Intervalsconcept:confidence-intervals
Confidence Intervals in the context of production machine-learning systems.
Continuous Batchingconcept:continuous-batching
Continuous Batching in the context of production machine-learning systems.
Contrastive Learningconcept:contrastive-learning
Contrastive Learning in the context of production machine-learning systems.
Convolutionconcept:convolution
Convolution in the context of production machine-learning systems.
Cosine Similarityconcept:cosine-similarity
Cosine Similarity in the context of production machine-learning systems.
Cost Thresholdsconcept:cost-thresholds
Cost Thresholds in the context of production machine-learning systems.
Coverageconcept:coverage
Coverage in the context of production machine-learning systems.
Cross Entropyconcept:cross-entropy
Cross Entropy in the context of production machine-learning systems.
Dbscanconcept:dbscan
Dbscan in the context of production machine-learning systems.
Dead Activationsconcept:dead-activations
Dead Activations in the context of production machine-learning systems.
Decision Utilityconcept:decision-utility
Decision Utility in the context of production machine-learning systems.
Decontaminationconcept:decontamination
Decontamination in the context of production machine-learning systems.
Delayed Labelsconcept:delayed-labels
Delayed Labels in the context of production machine-learning systems.
Dependency Locksconcept:dependency-locks
Dependency Locks in the context of production machine-learning systems.
Differential Privacyconcept:differential-privacy
Differential Privacy in the context of production machine-learning systems.
Direct Vs Recursiveconcept:direct-vs-recursive
Direct Vs Recursive in the context of production machine-learning systems.
Discrepancy Reportsconcept:discrepancy-reports
Discrepancy Reports in the context of production machine-learning systems.
Distance Concentrationconcept:distance-concentration
Distance Concentration in the context of production machine-learning systems.
Distributed Samplersconcept:distributed-samplers
Distributed Samplers in the context of production machine-learning systems.
Diversityconcept:diversity
Diversity in the context of production machine-learning systems.
Dpo Rlhfconcept:dpo-rlhf
Dpo Rlhf in the context of production machine-learning systems.
Driftconcept:drift
Drift in the context of production machine-learning systems.
Dropoutconcept:dropout
Dropout in the context of production machine-learning systems.
Embeddingsconcept:embeddings
Embeddings in the context of production machine-learning systems.
Empirical Riskconcept:empirical-risk
Empirical Risk in the context of production machine-learning systems.
Equalized Oddsconcept:equalized-odds
Equalized Odds in the context of production machine-learning systems.
Evaluationconcept:evaluation
Evaluation in the context of production machine-learning systems.
Event Timeconcept:event-time
Event Time in the context of production machine-learning systems.
Experiment Guardrailsconcept:experiment-guardrails
Experiment Guardrails in the context of production machine-learning systems.
Explained Varianceconcept:explained-variance
Explained Variance in the context of production machine-learning systems.
Exposure Biasconcept:exposure-bias
Exposure Bias in the context of production machine-learning systems.
Fairnessconcept:fairness
Fairness in the context of production machine-learning systems.
Fallbackconcept:fallback
Fallback in the context of production machine-learning systems.
Feature Scalingconcept:feature-scaling
Feature Scaling in the context of production machine-learning systems.
Feature Storesconcept:feature-stores
Feature Stores in the context of production machine-learning systems.
Feature Subsamplingconcept:feature-subsampling
Feature Subsampling in the context of production machine-learning systems.
Flink Sparkconcept:flink-spark
Flink Spark in the context of production machine-learning systems.
Focal Lossconcept:focal-loss
Focal Loss in the context of production machine-learning systems.
Forecast Originconcept:forecast-origin
Forecast Origin in the context of production machine-learning systems.
Forecast Scaleconcept:forecast-scale
Forecast Scale in the context of production machine-learning systems.
Fp16 Bf16concept:fp16-bf16
Fp16 Bf16 in the context of production machine-learning systems.
Fsdpconcept:fsdp
Fsdp in the context of production machine-learning systems.
Geluconcept:gelu
Gelu in the context of production machine-learning systems.
Global Batchconcept:global-batch
Global Batch in the context of production machine-learning systems.
Golden Inferenceconcept:golden-inference
Golden Inference in the context of production machine-learning systems.
Golden Intermediatesconcept:golden-intermediates
Golden Intermediates in the context of production machine-learning systems.
Golden Pathsconcept:golden-paths
Golden Paths in the context of production machine-learning systems.
Gqa Mqaconcept:gqa-mqa
Gqa Mqa in the context of production machine-learning systems.
Gradient Checksconcept:gradient-checks
Gradient Checks in the context of production machine-learning systems.
Gradient Clippingconcept:gradient-clipping
Gradient Clipping in the context of production machine-learning systems.
Gradient Flowconcept:gradient-flow
Gradient Flow in the context of production machine-learning systems.
Group Leakageconcept:group-leakage
Group Leakage in the context of production machine-learning systems.
Guardrailsconcept:guardrails
Guardrails in the context of production machine-learning systems.
Hbm Bandwidthconcept:hbm-bandwidth
Hbm Bandwidth in the context of production machine-learning systems.
Hidden Stateconcept:hidden-state
Hidden State in the context of production machine-learning systems.
Hinge Lossconcept:hinge-loss
Hinge Loss in the context of production machine-learning systems.
Hnswconcept:hnsw
Hnsw in the context of production machine-learning systems.
Idempotencyconcept:idempotency
Idempotency in the context of production machine-learning systems.
Immutable Data Snapshotsconcept:immutable-data-snapshots
Immutable Data Snapshots in the context of production machine-learning systems.
Implicit Feedbackconcept:implicit-feedback
Implicit Feedback in the context of production machine-learning systems.
Impurity Gainconcept:impurity-gain
Impurity Gain in the context of production machine-learning systems.
Initializationconcept:initialization
Initialization in the context of production machine-learning systems.
Intended Useconcept:intended-use
Intended Use in the context of production machine-learning systems.
Intersectional Slicesconcept:intersectional-slices
Intersectional Slices in the context of production machine-learning systems.
Iou Mapconcept:iou-map
Iou Map in the context of production machine-learning systems.
K Meansconcept:k-means
K Means in the context of production machine-learning systems.
Kafkaconcept:kafka
Kafka in the context of production machine-learning systems.
Kernelsconcept:kernels
Kernels in the context of production machine-learning systems.
KV Cacheconcept:kv-cache
KV Cache in the context of production machine-learning systems.
KV Layoutconcept:kv-layout
KV Layout in the context of production machine-learning systems.
Label Maturityconcept:label-maturity
Label Maturity in the context of production machine-learning systems.
Lag Featuresconcept:lag-features
Lag Features in the context of production machine-learning systems.
Latency Baselinesconcept:latency-baselines
Latency Baselines in the context of production machine-learning systems.
Lineageconcept:lineage
Lineage in the context of production machine-learning systems.
Log Oddsconcept:log-odds
Log Odds in the context of production machine-learning systems.
Loraconcept:lora
Lora in the context of production machine-learning systems.
Loss Scalingconcept:loss-scaling
Loss Scaling in the context of production machine-learning systems.
LSTM Gatesconcept:lstm-gates
LSTM Gates in the context of production machine-learning systems.
Metric Contractsconcept:metric-contracts
Metric Contracts in the context of production machine-learning systems.
MLflowconcept:mlflow
MLflow in the context of production machine-learning systems.
Model Supply Chainconcept:model-supply-chain
Model Supply Chain in the context of production machine-learning systems.
Moeconcept:moe
Moe in the context of production machine-learning systems.
Multi Head Attentionconcept:multi-head-attention
Multi Head Attention in the context of production machine-learning systems.
Multi Objective Trainingconcept:multi-objective-training
Multi Objective Training in the context of production machine-learning systems.
Nchw Tensorsconcept:nchw-tensors
Nchw Tensors in the context of production machine-learning systems.
Ndcgconcept:ndcg
Ndcg in the context of production machine-learning systems.
Next Token Lossconcept:next-token-loss
Next Token Loss in the context of production machine-learning systems.
Normalizationconcept:normalization
Normalization in the context of production machine-learning systems.
Output Activationsconcept:output-activations
Output Activations in the context of production machine-learning systems.
Padding Masksconcept:padding-masks
Padding Masks in the context of production machine-learning systems.
Paged Attentionconcept:paged-attention
Paged Attention in the context of production machine-learning systems.
Paper To Production Gapsconcept:paper-to-production-gaps
Paper To Production Gaps in the context of production machine-learning systems.
Parquet Lakehouseconcept:parquet-lakehouse
Parquet Lakehouse in the context of production machine-learning systems.
PCAconcept:pca
PCA in the context of production machine-learning systems.
Pinned Memoryconcept:pinned-memory
Pinned Memory in the context of production machine-learning systems.
Pr Curvesconcept:pr-curves
Pr Curves in the context of production machine-learning systems.
Pre Normconcept:pre-norm
Pre Norm in the context of production machine-learning systems.
Prediction Contractsconcept:prediction-contracts
Prediction Contracts in the context of production machine-learning systems.
Prediction Cutoffconcept:prediction-cutoff
Prediction Cutoff in the context of production machine-learning systems.
Prefill Decodeconcept:prefill-decode
Prefill Decode in the context of production machine-learning systems.
Prevalence Sensitivityconcept:prevalence-sensitivity
Prevalence Sensitivity in the context of production machine-learning systems.
Privacyconcept:privacy
Privacy in the context of production machine-learning systems.
Privacy Threat Modelingconcept:privacy-threat-modeling
Privacy Threat Modeling in the context of production machine-learning systems.
Profilersconcept:profilers
Profilers in the context of production machine-learning systems.
Prompt Injectionconcept:prompt-injection
Prompt Injection in the context of production machine-learning systems.
Proper Lossesconcept:proper-losses
Proper Losses in the context of production machine-learning systems.
Proper Scoring Rulesconcept:proper-scoring-rules
Proper Scoring Rules in the context of production machine-learning systems.
Provenanceconcept:provenance
Provenance in the context of production machine-learning systems.
Pseudo Residualsconcept:pseudo-residuals
Pseudo Residuals in the context of production machine-learning systems.
Qkv Projectionsconcept:qkv-projections
Qkv Projections in the context of production machine-learning systems.
Quantilesconcept:quantiles
Quantiles in the context of production machine-learning systems.
Quantizationconcept:quantization
Quantization in the context of production machine-learning systems.
Queueingconcept:queueing
Queueing in the context of production machine-learning systems.
RAGconcept:rag
RAG in the context of production machine-learning systems.
RAG Vs Tuningconcept:rag-vs-tuning
RAG Vs Tuning in the context of production machine-learning systems.
Randomization Unitsconcept:randomization-units
Randomization Units in the context of production machine-learning systems.
Rank World Sizeconcept:rank-world-size
Rank World Size in the context of production machine-learning systems.
Ranking Metricsconcept:ranking-metrics
Ranking Metrics in the context of production machine-learning systems.
Readinessconcept:readiness
Readiness in the context of production machine-learning systems.
Recall Latency Tradeoffconcept:recall-latency-tradeoff
Recall Latency Tradeoff in the context of production machine-learning systems.
Receptive Fieldsconcept:receptive-fields
Receptive Fields in the context of production machine-learning systems.
Reconciliationconcept:reconciliation
Reconciliation in the context of production machine-learning systems.
Recourseconcept:recourse
Recourse in the context of production machine-learning systems.
Regularizationconcept:regularization
Regularization in the context of production machine-learning systems.
Release Gatesconcept:release-gates
Release Gates in the context of production machine-learning systems.
Reluconcept:relu
Relu in the context of production machine-learning systems.
Relu Geluconcept:relu-gelu
Relu Gelu in the context of production machine-learning systems.
Replay Testsconcept:replay-tests
Replay Tests in the context of production machine-learning systems.
Reproducibilityconcept:reproducibility
Reproducibility in the context of production machine-learning systems.
Reproduction Laddersconcept:reproduction-ladders
Reproduction Ladders in the context of production machine-learning systems.
Rest Grpcconcept:rest-grpc
Rest Grpc in the context of production machine-learning systems.
Robust Regressionconcept:robust-regression
Robust Regression in the context of production machine-learning systems.
Rolling Backtestsconcept:rolling-backtests
Rolling Backtests in the context of production machine-learning systems.
Ropeconcept:rope
Rope in the context of production machine-learning systems.
Run Manifestsconcept:run-manifests
Run Manifests in the context of production machine-learning systems.
Samplingconcept:sampling
Sampling in the context of production machine-learning systems.
Schema Contractsconcept:schema-contracts
Schema Contracts in the context of production machine-learning systems.
Seasonal Naiveconcept:seasonal-naive
Seasonal Naive in the context of production machine-learning systems.
Second Order Boostingconcept:second-order-boosting
Second Order Boosting in the context of production machine-learning systems.
Seed Varianceconcept:seed-variance
Seed Variance in the context of production machine-learning systems.
Semantic Contractsconcept:semantic-contracts
Semantic Contracts in the context of production machine-learning systems.
Sftconcept:sft
Sft in the context of production machine-learning systems.
Shapconcept:shap
Shap in the context of production machine-learning systems.
Sigmoid Saturationconcept:sigmoid-saturation
Sigmoid Saturation in the context of production machine-learning systems.
Siluconcept:silu
Silu in the context of production machine-learning systems.
Slice Reliabilityconcept:slice-reliability
Slice Reliability in the context of production machine-learning systems.
Slis Slosconcept:slis-slos
Slis Slos in the context of production machine-learning systems.
Speculative Decodingconcept:speculative-decoding
Speculative Decoding in the context of production machine-learning systems.
Srmconcept:srm
Srm in the context of production machine-learning systems.
State Ttlconcept:state-ttl
State Ttl in the context of production machine-learning systems.
Stragglersconcept:stragglers
Stragglers in the context of production machine-learning systems.
Surrogate Objectivesconcept:surrogate-objectives
Surrogate Objectives in the context of production machine-learning systems.
Tanh Saturationconcept:tanh-saturation
Tanh Saturation in the context of production machine-learning systems.
Teacher Forcingconcept:teacher-forcing
Teacher Forcing in the context of production machine-learning systems.
Temporal Splitsconcept:temporal-splits
Temporal Splits in the context of production machine-learning systems.
Tensor Parallelismconcept:tensor-parallelism
Tensor Parallelism in the context of production machine-learning systems.
Tensor Shape Contractsconcept:tensor-shape-contracts
Tensor Shape Contracts in the context of production machine-learning systems.
Three Pass Readingconcept:three-pass-reading
Three Pass Reading in the context of production machine-learning systems.
Timeoutsconcept:timeouts
Timeouts in the context of production machine-learning systems.
Tokenizationconcept:tokenization
Tokenization in the context of production machine-learning systems.
Tokenizer Fertilityconcept:tokenizer-fertility
Tokenizer Fertility in the context of production machine-learning systems.
Transfer Learningconcept:transfer-learning
Transfer Learning in the context of production machine-learning systems.
Transform Stateconcept:transform-state
Transform State in the context of production machine-learning systems.
Two Tower Embeddingsconcept:two-tower-embeddings
Two Tower Embeddings in the context of production machine-learning systems.
Uncertaintyconcept:uncertainty
Uncertainty in the context of production machine-learning systems.
Watermarksconcept:watermarks
Watermarks in the context of production machine-learning systems.
Driftfailure:drift
Drift in the context of production machine-learning systems.
Feedback Loopfailure:feedback-loop
Feedback Loop in the context of production machine-learning systems.
Latencyfailure:latency
Latency in the context of production machine-learning systems.
Oomfailure:oom
Oom in the context of production machine-learning systems.
Prompt Injectionfailure:prompt-injection
Prompt Injection in the context of production machine-learning systems.
Supply Chainfailure:supply-chain
Supply Chain in the context of production machine-learning systems.
Training–serving skewfailure:training-serving-skew
A semantic or computational mismatch between features seen during training and inference.
Datalifecycle:data
Data in the context of production machine-learning systems.
Evaluationlifecycle:evaluation
Evaluation in the context of production machine-learning systems.
Inferencelifecycle:inference
Inference in the context of production machine-learning systems.
Monitoringlifecycle:monitoring
Monitoring in the context of production machine-learning systems.
Traininglifecycle:training
The repeatable path from validated data to a versioned model artifact.
Imagemodality:image
Spatial raster data used in classification, detection, segmentation, and multimodal systems.
Tabularmodality:tabular
Tabular in the context of production machine-learning systems.
Textmodality:text
Text in the context of production machine-learning systems.
Time Seriesmodality:time-series
Time Series in the context of production machine-learning systems.
ML platformrole:platform
Infrastructure and paved roads that let teams train, ship, observe, and govern models safely.
Reliabilityrole:reliability
Reliability in the context of production machine-learning systems.
CUDAtool:cuda
CUDA in the context of production machine-learning systems.
Kafkatool:kafka
Kafka in the context of production machine-learning systems.
MLflowtool:mlflow
MLflow in the context of production machine-learning systems.
PyTorchtool:pytorch
A tensor, autograd, neural-network, and distributed-training runtime.
Tritontool:triton
Triton in the context of production machine-learning systems.
vLLMtool:vllm
vLLM in the context of production machine-learning systems.
NLP / LLMtrack:nlp-llm
A focused route through sequence models, transformers, adaptation, evaluation, and language-model serving.
Recommendation Searchtrack:recommendation-search
Recommendation Search in the context of production machine-learning systems.
Vision Multimodaltrack:vision-multimodal
Vision Multimodal in the context of production machine-learning systems.
Case-derived patterns

Reusable, but never context-free.

Uber

A useful ML platform is a connected lifecycle, not a pile of independent tools.

Airbnb

The feature definition is the product; stores and compute engines are implementations.

DoorDash

Online feature serving is a latency-critical database product with ML semantics.

Spotify

Platform adoption is a product and socio-technical problem.

Netflix

The best abstraction preserves the user’s reasoning model while adding production guarantees.

Uber

Distributed training performance is the overlap of compute, communication, input, and synchronization.

OpenAI

At scale, metadata and coordination become a workload of their own.

NVIDIA

Parallelism is a placement problem constrained by bytes, shapes, and topology.

PyTorch

A checkpoint is useful only when it is committed, portable, and routinely restored.

PyTorch

Compiler wins are workload-specific artifacts that need the same release discipline as models.

Shopify

Online inference is a user-facing distributed system with a model inside.

Pinterest

Accelerators pay off only when the entire request path feeds them efficiently.

Pinterest

Batch inference is a bounded dataflow whose correctness includes exactly-once publication.

Meta

Distributed inference capacity is constrained by the busiest relevant slice, not the average box.

LinkedIn

A GenAI platform manages evidence and behavior across many versioned layers.

Uber

Drift detection is triage; impact and diagnosis decide the action.

Cloudflare

Generated data can behave like executable configuration and needs release engineering.

Cloudflare

Adversarial ML requires joined telemetry and active probes, not passive averages.

Uber

ML deployment safety is progressive evidence plus rapid reversibility.

Google

Feedback data identifies policy-conditioned behavior; correction needs overlap and controlled variance.

Twitter

Sometimes product redesign and user agency are stronger mitigations than model tuning.

LinkedIn

Fairness infrastructure must be context-aware, governed, and integrated into decisions.

Google

Privacy is an end-to-end guarantee with explicit adversary, accounting, and utility tradeoffs.

Stripe

Risk modeling connects delayed supervision, adversarial drift, calibrated decisions, and layered controls.