Which AI model should you actually use?

Still the majority of the machine learning that actually runs in production at real companies.

• Linear, ridge, lasso, elastic netgives you a predicted number

  Draws the straight-line relationship between your inputs and a number, and reads a prediction off it.

  Use when Small to medium data with a roughly straight-line relationship, when you need to point to exactly which factor drove the number.

• Logistic regressiongives you a category

  Weighs the factors and outputs the probability that something is true or false.

  Use when The honest baseline for any yes-or-no question, and the standard anywhere a decision must be justified one factor at a time.

• Decision tree (CART)gives you a category

  Learns a flowchart of yes-or-no questions that a person can read and follow by hand.

  Use when You want plain if-this-then-that rules a human can execute, audit, or argue with.

• Random forestgives you a category

  Asks hundreds of slightly different decision trees and takes their majority vote.

  Use when A strong, low-maintenance default that copes with messy, mixed, and missing data without much tuning.

• Gradient-boosted trees: XGBoost, LightGBM, CatBoostgives you a predicted number

  Builds decision trees one after another, each fixing the mistakes of the last.

  Use when The default winner on spreadsheet-shaped data. Reach for this first unless you have a specific reason not to.

• Generalized additive models (GAM, EBM)gives you a predicted number

  Captures curved relationships while still showing a readable graph of how each factor moves the answer.

  Use when You need near-top accuracy and a regulator or clinician must see exactly how each variable affects the result.

• Naive Bayesgives you a category

  Counts how often each clue goes with each outcome and multiplies the odds together.

  Use when A very fast, very cheap baseline, especially on text and small data.

• k-nearest neighborsgives you a category

  Predicts by finding the most similar past examples and copying their answer.

  Use when Small datasets, or as a quick lookup over a set of similarity fingerprints.

• Support vector machine (SVM, SVR)gives you a category

  Draws the widest possible dividing line between two groups.

  Use when High-dimensional data with very few examples, like text or lab spectra.

• Gaussian processesgives you a predicted number

  Predicts a value and, just as importantly, how sure it is about that value.

  Use when Data is scarce or expensive to collect and you need honest confidence ranges, not just a guess.

• Bayesian networksgives you found and organized information

  A map of which things cause or depend on which, used to reason about probabilities.

  Use when You have expert knowledge of how factors relate and need to reason under uncertainty.

• AutoML: AutoGluon, H2O, FLAMLgives you a category

  Automatically tries many models and settings and hands you the best one.

  Use when Day one of a spreadsheet project, to set a strong benchmark before you invest in anything custom.

No labels. You are asking the data what shape it already has.

• K-meansgives you a group

  Sorts records into a set number of groups by how close together they are.

  Use when Fast, simple grouping when you can reasonably guess how many groups to expect.

• DBSCAN and HDBSCANgives you a group

  Finds groups as dense clumps of points and leaves the stragglers as outliers.

  Use when Groups of odd shapes, an unknown number of groups, and you want outliers flagged too.

• Gaussian mixture modelgives you a group

  Groups records but lets each one partly belong to several groups at once.

  Use when Overlapping groups where you want a probability of membership, not a hard label.

• Hierarchical clusteringgives you a group

  Builds a family tree of similarity, from every record alone up to one big group.

  Use when You want the whole tree of relationships, not a single fixed set of groups.

• PCA and SVDgives you found and organized information

  Squeezes many columns down to a few that still carry most of the information.

  Use when Compress, de-noise, and simplify data before another model uses it.

• Non-negative matrix factorizationgives you found and organized information

  Breaks data into a handful of additive building blocks you can interpret.

  Use when You want parts that add up to the whole, like topics in text or ingredients in a signal.

• Independent component analysisgives you found and organized information

  Untangles several signals that got mixed together into one recording.

  Use when Separating mixed sources, such as isolating brain signals from noise.

• t-SNE and UMAPgives you found and organized information

  Flattens complex data onto a 2-D map so you can see clusters with your eyes.

  Use when Visualization and exploration only. Never feed its output into another model.

• Latent Dirichlet allocationgives you a group

  The classic way to discover the hidden topics running through a pile of documents.

  Use when A large text collection where you want readable topics and have no GPU.

• BERTopicgives you a group

  Groups documents by meaning and auto-labels each cluster with its keywords.

  Use when Modern topic discovery. Better than the classic approach in almost every case.

• Association rules: Apriori, FP-Growthgives you found and organized information

  Finds patterns like people who bought A also bought B.

  Use when Shopping baskets and co-occurrence, when you want explicit rules you can read.

• Isolation Forest, one-class SVM, LOFgives you a category

  Learns what normal looks like and flags whatever does not fit, with no examples of bad.

  Use when Catching the unusual when you have no labelled examples of fraud or faults, which is most of the time.

• Autoencoder for anomaly detectiongives you a category

  Learns to rebuild normal data; whatever it rebuilds badly is the anomaly.

  Use when Spotting the odd one out when normal behaviour is complex and high-dimensional.

• Self-organizing mapgives you a group

  Arranges data on a 2-D grid so similar things sit near each other.

  Use when A niche, older tool for visual clustering in industrial monitoring.

A completely separate discipline from language models, and one of the highest-return areas in a real business.

• ARIMA, SARIMA, SARIMAXgives you a predicted number

  Forecasts by learning a series' own momentum, its seasonal pattern, and outside drivers.

  Use when A handful of series, a stable pattern, and you need statistical rigour and real confidence intervals.

• Exponential smoothing, Holt-Winters, ETSgives you a predicted number

  Forecasts from a weighted average of recent history, trend, and season.

  Use when A fast, sturdy baseline. Start here before anything more complex.

• Prophetgives you a predicted number

  A forecaster built for analysts that handles trends, seasons, and holidays out of the box.

  Use when Business forecasts with known holiday and promotion spikes.

• GARCHgives you a predicted number

  Forecasts how much a number will swing, not where it will land.

  Use when Financial risk and any series where the size of the ups and downs is what matters.

• Kalman filter and state-space modelsgives you a predicted number

  Continuously updates its best estimate of a hidden value as noisy new readings arrive.

  Use when Live tracking from noisy sensors, like position from GPS.

• Gradient boosting on lagged featuresgives you a predicted number

  Turns forecasting into a spreadsheet problem by feeding past values in as columns.

  Use when Many series with lots of extra context. The quiet winner of most real forecasting contests.

• DeepARgives you a predicted number

  Learns one forecasting model across thousands of related series at once, with ranges.

  Use when Hundreds or thousands of related series where you want full probability ranges.

• N-BEATS and N-HiTSgives you a predicted number

  Deep-learning forecasters that need no manual feature work.

  Use when Strong single-series accuracy with minimal setup.

• Temporal fusion transformergives you a predicted number

  A forecaster that weighs many inputs and shows which ones drove the prediction.

  Use when Complex forecasts with many drivers where you must explain what moved the number.

• PatchTST and iTransformergives you a predicted number

  Transformer forecasters tuned for very long histories and many series.

  Use when Long-horizon, many-variable forecasting like grid load or sensor fleets.

• Foundation models: TimesFM, Chronos, Moiraigives you a predicted number

  Pre-trained forecasters that work on a brand-new series with no training.

  Use when You need a fast forecast on a new series and have no time to train one.

• Survival analysis: Kaplan-Meier, Cox, DeepSurvgives you a predicted number

  Predicts how long until an event, correctly handling cases that have not happened yet.

  Use when The question is when, not whether: churn timing, equipment failure, contract renewal.

• Change-point detection: PELT, BOCPDgives you a category

  Spots the moment a trend genuinely shifts, separate from normal noise.

  Use when Deciding whether a metric really changed on a given date, or you are chasing noise.

The primitives. Most of the named products in this guide are one of these wearing a costume.

• Multilayer perceptrongives you a predicted number

  The basic neural network: layers of simple units that together learn any pattern.

  Use when The final decision layer sitting on top of a bigger model. Rarely the whole answer by itself.

• Convolutional networks: ResNet, EfficientNet, ConvNeXtgives you a category

  A neural network that scans an image in small patches to recognize shapes and textures.

  Use when Image tasks that must be small, fast, or run on a device. Still the go-to for everyday vision.

• RNN, LSTM, GRUgives you a predicted number

  A network that reads a sequence one step at a time, carrying a memory of what came before.

  Use when Sequences on tiny hardware or under strict live-streaming limits.

• Transformergives you found and organized information

  The architecture behind modern AI: it weighs every part of the input against every other part.

  Use when An architecture, not a product. Vision models, speech models, forecasters, and AlphaFold are all transformers, and none of them are language models.

• Vision transformergives you a category

  A transformer that treats an image as a grid of patches instead of words.

  Use when Large image datasets where you want strong transfer learning. Needs more data than a CNN.

• Graph neural networks: GCN, GraphSAGE, GATgives you a category

  A network that learns from how things are connected, passing information along the links.

  Use when Your data is a network of relationships and the connections carry the signal.

• Autoencoder and VAEgives you found and organized information

  Squeezes data down to a compact code and rebuilds it, learning the essence in between.

  Use when De-noising, anomaly detection, and compression. The compact code is the useful part.

• Generative adversarial network (GAN)gives you newly generated content

  Two networks compete: one forges, one detects, until the forgeries look real.

  Use when Fast generation, realistic fake tabular data, and image upscaling. Largely replaced by diffusion for images.

• Diffusion modelsgives you newly generated content

  Starts from pure noise and repeatedly cleans it up into a realistic image or signal.

  Use when The highest-quality generation across images, video, audio, and molecules.

• Flow matching and rectified flowgives you newly generated content

  A newer generator that learns a direct path from noise to a finished result.

  Use when Modern image generation that reaches quality in far fewer steps than diffusion.

• Normalizing flowsgives you a predicted number

  A generator that can also state the exact likelihood of any given sample.

  Use when You need an exact probability of a sample, not just the ability to produce one.

• State-space models: Mamba, S4gives you found and organized information

  Processes very long sequences efficiently, in one pass, without attention's cost.

  Use when Extremely long inputs, live streaming, and tight efficiency budgets.

• Mixture of expertsgives you found and organized information

  Routes each piece of input to a few specialist sub-networks instead of the whole model.

  Use when Growing a model's capacity without growing the cost of each prediction. Most frontier LLMs work this way.

• NeRF and 3D Gaussian splattinggives you newly generated content

  Turns a handful of ordinary photos into an explorable 3-D scene.

  Use when Digital twins, as-built capture, property walkthroughs, and visual effects.

• Physics-informed neural networksgives you a predicted number

  A network that is forced to obey known physical laws while it learns.

  Use when Sparse data plus known physics: fluid flow, structures, subsurface modelling.

• Siamese and contrastive networksgives you a category

  Learns to tell whether two things are the same or different.

  Use when Verification and similarity when you have very few examples of each item, like matching signatures.

Organized by task, not by model. This is the part of AI that has nothing to do with language at all.

• Image classification: EfficientNet, ConvNeXt, ViTgives you a category

  Gives a whole image one label.

  Use when The entire image gets a single verdict, like pass or fail.

• Object detection: YOLO, RT-DETR, Grounding DINOgives you found and organized information

  Draws a box around each object and says what it is.

  Use when You need to know what is in the frame, where, and how many.

• Segmentation: U-Net, Mask R-CNN, SAM 2gives you found and organized information

  Traces the exact outline of each object, down to the pixel.

  Use when You need the precise shape or area, not just that something is present.

• Pose and keypoint estimation: MediaPipe, RTMPosegives you found and organized information

  Tracks the position of body, hand, and joint points.

  Use when Movement and posture are the signal: ergonomics, physical therapy, sports.

• Face detection and recognition: RetinaFace, ArcFacegives you a category

  Finds faces and matches them to identities.

  Use when Access control and identity. Gate this hard on consent and law.

• OCR: Tesseract, PaddleOCR, Surya, TrOCRgives you found and organized information

  Reads printed text out of an image into characters.

  Use when Any time text is trapped inside a scan or photo.

• Document AI: LayoutLMv3, Donut, Nougatgives you found and organized information

  Reads a document and understands which text is which field, plus tables.

  Use when You need clean structured data out of forms and invoices, not a wall of text.

• Handwriting recognition: TrOCR, Krakengives you found and organized information

  Reads handwritten and cursive text.

  Use when Historical records, clinical notes, and handwritten field forms.

• Object tracking: ByteTrack, DeepSORTgives you found and organized information

  Follows the same object from frame to frame across a video.

  Use when Queue times, dwell time, throughput, and sports analytics.

• Optical flow: RAFTgives you found and organized information

  Measures how each pixel moves between video frames.

  Use when When motion itself is what you are measuring.

• Depth estimation: Depth Anything, MiDaSgives you found and organized information

  Estimates how far away things are from a single flat photo.

  Use when Robotics, AR, and taking measurements from an ordinary picture.

• Visual anomaly detection: PatchCore, PaDiMgives you a category

  Learns what a good part looks like and flags anything different.

  Use when Factory quality control, where you have good samples but not every possible defect.

• Super-resolution and restoration: Real-ESRGAN, SwinIRgives you newly generated content

  Sharpens, upscales, and repairs degraded images.

  Use when The source is low-quality and the detail matters.

• Zero-shot vision: CLIP, SigLIPgives you a category

  Finds or labels images from a plain text description, with no training.

  Use when A huge image library with no labels: find every photo with a forklift.

• Image generation: Stable Diffusion, Flux, Imagen, Ideogramgives you newly generated content

  Creates a new image from a text description.

  Use when You need a picture that does not exist. Ideogram when the text inside the image must be legible.

• Conditioning and control: ControlNet, LoRA, IP-Adaptergives you newly generated content

  Steers image generation to hold a pose, layout, or brand style.

  Use when Generation alone is not enough; you need it on-brand and on-composition.

• Video generation: Sora, Veo, Runway, Klinggives you newly generated content

  Creates moving video from a text prompt or a still image.

  Use when You need footage you cannot practically shoot.

• Video understanding: VideoMAE, TimeSformergives you a category

  Recognizes what is happening across a stretch of video.

  Use when The event unfolds over time; a single frame will not tell you.

• Multimodal LLMgives you found and organized information

  A language model that can also look at an image and reason about it.

  Use when The question about an image is open-ended and needs judgement, not a fixed label.

An entire modality with nothing to do with text generation.

• Speech recognition: Whisper, Parakeet, Conformergives you found and organized information

  Turns spoken words into written text.

  Use when Any spoken input: calls, meetings, voice notes.

• Speaker diarization: pyannotegives you found and organized information

  Works out who spoke and when.

  Use when More than one person is talking and you need it attributed.

• Speaker verification: ECAPA-TDNNgives you a category

  Confirms whether a voice belongs to a specific person.

  Use when Verifying identity from a voice.

• Text to speech: ElevenLabs, Kokoro, Pipergives you newly generated content

  Reads written text aloud in a natural voice.

  Use when The output needs to be heard, not read: phone systems, accessibility, narration.

• Voice conversion: RVC, So-VITSgives you newly generated content

  Changes whose voice a recording sounds like.

  Use when Dubbing and localization. Consent and law gate this, not the technology.

• Audio event classification: YAMNet, PANNs, ASTgives you a category

  Recognizes non-speech sounds.

  Use when The signal is a sound, not a word: a failing bearing, breaking glass, an alarm.

• Music generation: MusicGen, Stable Audiogives you newly generated content

  Generates music and sound effects.

  Use when You need original audio or a background track.

• Source separation: Demucs, Spleetergives you newly generated content

  Splits a mixed recording into separate tracks.

  Use when You need one element, like a single voice, out of a noisy mix.

• Keyword spottinggives you a category

  Listens continuously for a specific wake word using almost no power.

  Use when Always-on, battery-powered listening for a wake word.

• Audio embeddings: CLAP, VGGishgives you found and organized information

  Turns sounds into searchable fingerprints.

  Use when Searching or grouping audio by how it sounds, like every clip of a failing bearing.

One family out of fourteen. The front door of the building, not the building.

• Frontier LLM: Claude, GPT-class, Geminigives you newly generated content

  The large, general models that write and reason over language.

  Use when Hard reasoning, agent workflows, and genuinely open-ended language.

• Reasoning models (extended thinking)gives you newly generated content

  A language model that deliberately thinks longer before it answers.

  Use when Multi-step maths, planning, proofs, and difficult code.

• Small language models: Haiku-class, Phi, Gemmagives you newly generated content

  Compact, cheaper language models that still handle everyday language tasks.

  Use when High volume, low latency, on-device, or cost-sensitive work.

• Code models: Claude Code, Qwen-Codergives you newly generated content

  Language models specialised in writing, fixing, and reviewing code.

  Use when The output is source code.

• Domain fine-tune (LoRA, SFT)gives you newly generated content

  A base model further trained on your own examples to specialise it.

  Use when A narrow, repeated task where you already have labelled examples.

• Encoder-decoder: T5, BART, Pegasusgives you newly generated content

  Language models built to transform one text into another, predictably.

  Use when Summarize, translate, or restructure text. A transform, not a conversation.

Cheaper, faster, and usually more accurate than a language model. The boring, profitable half of NLP.

• Text embeddings: BGE, E5, Voyage, Coheregives you found and organized information

  Turns text into number-fingerprints so a computer can measure meaning and similarity.

  Use when The engine under search, retrieval, clustering, de-duplication, and find-similar.

• Sparse retrieval: BM25, SPLADEgives you a ranking or recommendation

  Classic keyword search that ranks documents by exact word matches.

  Use when Part numbers, IDs, and names, exactly where meaning-based search quietly fails. Pair it with embeddings.

• Rerankers: Cohere Rerank, BGE-rerankergives you a ranking or recommendation

  Takes a shortlist of results and re-orders it for a much sharper top few.

  Use when The single highest-value upgrade to any search or retrieval system.

• Text classification: DeBERTa, DistilBERT, fastTextgives you a category

  Reads text and drops it into predefined categories, at scale and cheaply.

  Use when High-volume labelling: tickets, intent, sentiment, moderation. These understand text without generating it, at a fraction of an LLM's cost.

• Named entity recognition: spaCy, GLiNERgives you found and organized information

  Pulls the names, dates, and amounts out of free text.

  Use when You need specific fields extracted, not a summary.

• Relation and event extractiongives you found and organized information

  Finds how the entities in text relate, and links them up.

  Use when Turning documents into a connected knowledge graph.

• Machine translation: NLLB, MarianMT, SeamlessM4Tgives you newly generated content

  Translates text between languages.

  Use when Bulk translation, where paying an LLM per word would be ruinous.

• Extractive question answeringgives you found and organized information

  Answers a question by quoting the exact sentence from a document, inventing nothing.

  Use when Compliance and legal answers where a made-up answer is unacceptable.

• Safety and moderation classifiers: Llama Guardgives you a category

  Screens text and blocks unsafe or policy-violating content.

  Use when Any system open to the public or to untrusted input.

• Grammar and style: LanguageToolgives you a category

  Checks writing against fixed grammar and style rules.

  Use when You want consistency and correctness, not creativity.

Trained on your behavioral log, which a language model has never seen.

• Matrix factorization: ALS, SVD++gives you a ranking or recommendation

  Learns hidden taste patterns from who interacted with what.

  Use when You have an activity log and few details about the items themselves.

• Content-based filteringgives you a ranking or recommendation

  Recommends items similar to ones a person already liked, by their features.

  Use when New items with rich descriptions and little interaction history yet.

• Two-tower retrievalgives you a ranking or recommendation

  Instantly narrows millions of items to a relevant shortlist for each person.

  Use when Stage one of every modern recommender: a thousand candidates from millions, in milliseconds.

• Ranking models: Wide and Deep, DeepFM, DLRMgives you a ranking or recommendation

  Scores each shortlisted item by how likely the person is to click or buy.

  Use when Stage two: putting the shortlist in the best order.

• Sequential recommenders: SASRec, BERT4Recgives you a ranking or recommendation

  Predicts the next item from the order of what someone did before.

  Use when Session-based what-comes-next, like the next episode or track.

• Graph recommenders: LightGCN, PinSagegives you a ranking or recommendation

  Recommends by spreading signals across the web of users and items.

  Use when Sparse activity with strong word-of-mouth or network effects.

• Learning to rank: LambdaMARTgives you a ranking or recommendation

  Trains directly on the quality of the final ordering.

  Use when When getting the order of results right is the actual job, like site search.

Reinforcement learning is powerful and hard. Bandits are cheap, safe, and badly underused. Start with bandits.

• Multi-armed bandit: Thompson sampling, UCBgives you a decision or plan

  Tries several options and steadily shifts toward the winner while still testing the rest.

  Use when Live experiments where you want to stop wasting traffic on the losing option: headlines, layouts, offers.

• Contextual banditgives you a decision or plan

  Like a bandit, but it looks at who each person is, so the best option can differ per person.

  Use when Personalized choices with instant feedback, like which offer to show this particular visitor.

• Q-learning and DQNgives you a decision or plan

  Learns by trial and error which action pays off in each situation, the way you learn a game by playing it.

  Use when Step-by-step decision problems where you can practise safely in a simulator first.

• Policy gradient: PPO, SAC, TD3gives you a decision or plan

  Learns a strategy for smooth, continuous control like steering or throttle.

  Use when Robotics and control problems trained against a simulator.

• Model-based RL and MuZerogives you a decision or plan

  Builds its own model of how the world reacts, then plans ahead inside it.

  Use when When real trial and error is too slow or costly, so learning efficiency matters.

• Offline RL and decision transformersgives you a decision or plan

  Learns the best sequence of actions from a fixed history, never experimenting live.

  Use when When you cannot safely experiment on real customers or patients, so you learn from records.

• RLHF, DPO, GRPOgives you a decision or plan

  Tunes a generative model using human preferences about which answers are better.

  Use when Post-training a language model to be helpful, safe, and on brand.

Not AI in the marketing sense, and frequently the actual answer.

• Mixed-integer programming: OR-Tools, Gurobi, CPLEXgives you a decision or plan

  Finds the provably best plan that never breaks a hard rule.

  Use when Scheduling, routing, and budgets with constraints that cannot be violated.

• Constraint programming: CP-SATgives you a decision or plan

  Fits many rigid requirements together into a workable schedule.

  Use when Rostering, timetabling, and assignment puzzles.

• Vehicle routing solversgives you a decision or plan

  Plans the most efficient routes under delivery windows and vehicle limits.

  Use when Fleets, deliveries, and field-service dispatch.

• Genetic algorithms and simulated annealinggives you a decision or plan

  Searches a huge space of options by mutating and improving candidate solutions.

  Use when The goal is a black box or has no clean formula: layout, design, tuning.

• Bayesian optimizationgives you a decision or plan

  Chooses the next experiment to run so you learn the most from the fewest trials.

  Use when Each trial is expensive: lab work, physical tests, or model tuning.

• Monte Carlo simulationgives you a predicted number

  Rolls the dice thousands of times to show the full range of outcomes.

  Use when You need a risk range and odds, not a single point estimate.

• Discrete-event and agent-based simulationgives you a predicted number

  Builds a working model of a system so you can test changes before they are real.

  Use when Try a policy on a simulated warehouse, queue, or market before deploying it.

Predictive models tell you what will happen. These tell you what will happen if you act. That difference is worth a great deal of money.

• Randomized experiment (A/B test)gives you a predicted number

  Splits people at random to measure the true effect of a change.

  Use when Whenever you can genuinely randomize. This is the gold standard.

• Uplift and CATE: causal forest, T-learner, X-learnergives you a decision or plan

  Estimates how much a treatment changes each individual, not just the average.

  Use when Targeting the people an offer will actually sway, rather than those who would act anyway.

• Propensity score matchinggives you a predicted number

  Builds a fair comparison group from data you already have.

  Use when You could not run an experiment but still need a credible cause-and-effect claim.

• Difference in differencesgives you a predicted number

  Compares the before-and-after of a treated group against an untreated one.

  Use when A change rolled out to some regions or teams and not others.

• Synthetic controlgives you a predicted number

  Builds a stand-in for what would have happened from a blend of similar cases.

  Use when One treated unit, like a single store or market, and many untreated ones.

• Instrumental variablesgives you a predicted number

  Isolates a true effect when hidden factors are muddying the data.

  Use when Hidden confounding is distorting a plain comparison.

• Marketing mix modeling: Robyn, Meridiangives you a predicted number

  Estimates how much each channel actually contributes to sales.

  Use when Privacy-safe, top-down decisions about where to put the marketing budget.

• Structural causal models: DoWhy, EconMLgives you found and organized information

  Writes your cause-and-effect assumptions as an explicit, testable diagram.

  Use when High-stakes does-X-cause-Y claims you will have to defend.

Where AI stopped being a chatbot and started being an instrument.

• Protein structure: AlphaFold 3, ESMFold, Boltzgives you newly generated content

  Predicts a protein's 3-D shape from its sequence.

  Use when Structural biology and drug discovery.

• Molecular models: property GNNs, generative design, dockinggives you newly generated content

  Predicts how molecules behave and designs new ones.

  Use when Chemistry and pharmaceutical screening.

• Weather: GraphCast, Aurora, GenCastgives you a predicted number

  Forecasts weather as well as supercomputer models, far faster and cheaper.

  Use when Energy, agriculture, logistics, and insurance planning.

• Materials: GNoME, M3GNetgives you newly generated content

  Discovers new stable materials before anyone makes them.

  Use when Materials science and battery research.

• Geospatial: Prithvi, satellite segmentationgives you found and organized information

  Reads satellite and aerial imagery to map what is on the ground.

  Use when Land use, deforestation, crop yield, and damage assessment.

• Medical imaging: nnU-Net, MedSAMgives you found and organized information

  Finds and outlines structures in CT, MRI, and X-ray scans.

  Use when Clinical imaging and screening support.

• Genomics: Enformer, DNABERT, Evogives you a predicted number

  Reads DNA to predict what genes do and how variants matter.

  Use when Genomics research and interpreting variants.

• Robotics: RT-2, OpenVLA (vision-language-action)gives you a decision or plan

  Turns what a robot sees plus an instruction directly into movement.

  Use when Machines that must perceive and act in the physical world.

• Autonomous driving: BEVFormer, occupancy networksgives you a decision or plan

  Turns camera and sensor feeds into a driving picture and a plan.

  Use when Self-driving and driver-assistance systems.

Not models, but you cannot ship the models without them.

• Vector databases: pgvector, Qdrant, Pineconegives you a ranking or recommendation

  Stores those text fingerprints and finds the closest matches instantly.

  Use when The moment you use embeddings, they need somewhere to live and be searched.

• Feature stores: Feast, Tectongives you found and organized information

  Serves the exact same inputs to training and to the live model.

  Use when Prevents the most common silent production bug: training and serving seeing different data.

• Quantization and distillation: GPTQ, AWQ, GGUFgives you found and organized information

  Shrinks a model to run cheaper, faster, or on smaller hardware.

  Use when The model works but you cannot afford to run it as it stands.

• Synthetic data: CTGAN, SDVgives you newly generated content

  Generates realistic fake data that carries no private information.

  Use when Real data is restricted, or a rare case is badly under-represented.

• Explainability: SHAP, LIMEgives you found and organized information

  Shows which inputs drove a given prediction.

  Use when Required for regulated decisions, and your best debugging tool everywhere else.

• Drift monitoring: Evidently, NannyMLgives you a category

  Watches for when live data drifts away from what the model was trained on.

  Use when Models silently decay; you want to catch it before the business does.

• Guardrail modelsgives you a category

  Screen and block bad inputs and outputs at the edge of the system.

  Use when Any system exposed to untrusted input.