The prediction layer
for your AI agent.

JITM.ai (Just-In-Time Models) is an AutoML platform that turns any tabular dataset into a live prediction API. You upload a CSV, pick a column to predict, and within seconds you have a trained model with a production REST endpoint. It's designed to be used equally well by humans through a web UI and by AI agents through MCP.

No. JITM.ai automates feature engineering, algorithm selection, hyperparameter tuning, validation, and deployment. If you can identify which column you want to predict, the platform does the rest. Experienced ML practitioners still get deep telemetry: per-fold metrics, feature importance, leakage warnings, ensemble breakdowns, and downloadable holdout predictions.

Pair JITM.ai with your favourite LLM. Add https://mcp.jitm.ai to ChatGPT, Claude, Gemini, Cursor, or any MCP-compatible client, sign in, and you get a hands-on copilot that can upload your data, explain what the column analysis is showing, recommend which target to predict, kick off training, interpret the resulting metrics, and call your prediction endpoint on your behalf. You don't need to learn ML vocabulary to get started: ask things like "which column should I predict?", "is 0.84 accuracy good for this problem?", or "why is recall lower than precision?" and your LLM will translate everything into plain language while driving the platform through its 13 MCP tools. We think this is the fastest way to go from a CSV to a working model, especially if machine learning isn't your day job.

Anything you can frame as "given these inputs, predict this output" on tabular data. That covers classification (customer churn, fraud, medical diagnosis), regression (price, demand, lifetime value), and multi-class problems (product category, risk tier). The platform auto-detects the task type from your target column, and time-stamped data is handled with chronological splits automatically.

Those are libraries that still need code, infrastructure, and a data engineer to stand up. JITM.ai is the whole stack: ingestion, feature engineering, training, model selection, versioning, and an authenticated REST endpoint, all behind a single API call. The same capabilities are exposed through MCP, so an autonomous agent can put production ML into service without a human in the loop.

On Pro and Enterprise, Phase 2 trains a diverse portfolio of four families in parallel: XGBoost and LightGBM as the gradient-boosted workhorses, elastic-net GLM as a linear anchor that smooths predictions in sparse regions, and TabM as a parameter-efficient neural network with 32 internal ensemble heads that captures the smooth non-linear interactions the tree families can only approximate with step functions. CatBoost is added automatically when the dataset has the high-cardinality categorical mix it excels at. After training, three aggregation strategies compete on out-of-fold predictions (a CV-score-weighted blend, a stacked meta-learner, and a Caruana-style hill-climb selector), and whichever scores best on the chosen metric is the one that gets deployed. Diversity matters: all-gradient-boosted ensembles tend to fail in correlated ways at the tails of the distribution, while mixing in linear and neural families produces calibrated predictions across the full prediction range.

Training runs in two phases. Phase 1 is a fast randomised hyperparameter search that delivers a working baseline in seconds, sufficient for many use cases. Phase 2 (Builder tier and above) does a deeper, wider search on the full dataset with more trials, cross-validation, and multi-family ensembling on higher tiers. The platform automatically promotes the better version as the active inference endpoint while keeping Phase 1 available for comparison.

Phase 2 uses 5-fold cross-validation on Pro and Enterprise so accuracy is an average of multiple independent evaluations rather than one lucky split. Enterprise additionally trains each fold across multiple random seeds and averages the predictions, which damps out the variance that a single seed can introduce on smaller or noisier datasets. We also run active leakage detection: when training scores are dramatically better than validation and validation itself looks suspiciously good, we flag the likely causes (target-derived features, ID leakage, or future-into-past leakage on time-series). Metadata-flagged columns such as IDs and meta_* prefixes are excluded from features by default.

Yes, automatically. When JITM.ai detects a datetime column it switches to chronological validation: the holdout is always the most recent slice of data, and cross-validation uses forward-only folds so the model never trains on rows more recent than what it's predicting. We also strip monotonic date parts (like year, or a raw day index) from feature engineering on the time axis, since those create spurious distribution shifts that flatter training and break in production.

Missing numeric values are median-imputed, missing categoricals get most-frequent imputation, and both carry a missing-indicator flag so the model can learn that missingness itself is informative. Low-cardinality categoricals are one-hot encoded; high-cardinality ones are ordinal encoded. Datetime columns are expanded into seasonal features (month, day-of-week, hour, is-weekend). Group columns and meta-prefixed identifiers are auto-detected and kept out of the feature set.

Yes. Every trained model comes with global feature importance ranking which columns drive predictions overall. Pro and Enterprise tiers add SHAP-based explanations, which show per-feature contributions to each individual prediction. Both views are available in the dashboard and through the API.

CSV, JSON, and Parquet are all supported. Uploads are streamed through presigned URLs, so effective file size is bounded by tier quota rather than HTTP limits. On upload we run column type inference, missing-value analysis, quality checks, and target candidate detection, all before you commit to training.

Typical p99 latency is under 25 milliseconds. We achieve this with in-memory model caching, a single vectorised pass through the feature pipeline at prediction time, and pre-warming on server start. That's fast enough for real-time systems: trading engines, recommendation services, fraud scoring, and IoT control loops.

Every trained model gets a dedicated REST endpoint and its own per-model access token. Send a JSON payload of feature values; receive a prediction and, for classification, per-class probabilities. Account-level MCP tokens also work on the same endpoint, so humans, services, and agents all use the same interface.

Paste https://mcp.jitm.ai into any MCP-compatible client (Claude, Cursor, ChatGPT, Windsurf) and sign in with your JITM.ai account. That's it. OAuth handles authentication automatically, so there are no tokens to generate or config files to edit. The agent discovers all 13 tools on its own, and a single jitm_describe_workflow call teaches it the full pipeline. For CI pipelines or headless environments without a browser, Personal Access Tokens are still supported as a fallback.

Yes. Every dataset and model is scoped to your account and stored in encrypted object storage. Data is never shared between users and never used to train shared models. MCP agents authenticate via OAuth with short-lived sessions tied to your identity, so no long-lived secrets are stored in agent configs. For programmatic access, scoped Personal Access Tokens (per-model or account-level) can be rotated or revoked at any time.

Free gets Phase 1 XGBoost models for quick experiments. Builder adds Phase 2 deep training with a wider hyperparameter search on the full dataset. Pro and Enterprise unlock the multi-family ensemble (XGBoost, LightGBM, GLM, TabM, plus CatBoost when beneficial) combined via 5-fold cross-validation, an OOF tournament between weighted blending, a stacked meta-learner and Caruana hill-climb that picks the best aggregator on your data, and SHAP explainability. Enterprise raises the trial budget and trains each fold across three random seeds (vs two on Pro) for lower-variance predictions.