One governed data plane. Five specialized engines.
Every engine automates a stage of the data science lifecycle and shares the same lineage, governance, and security model. Adopt one, or run the full stack from ingestion to decision.
01 · Automated feature intelligence
Medovac Signal
Feature engineering that discovers, validates, and governs itself.
Medovac Signal is an automated feature intelligence engine. It connects directly to your lakehouse, profiles every column, and uses gradient-boosted importance scoring combined with mutual-information ranking to synthesize candidate features. Each candidate is stress-tested for leakage, drift, and multicollinearity before it is ever exposed to a model. Approved features land in a versioned, point-in-time-correct feature store with full lineage, so the same definition powers training, batch scoring, and low-latency online serving without skew.
Techniques
Key capabilities
Leakage-aware synthesis
Every generated feature is checked against target timelines using point-in-time joins, so future information never contaminates training sets.
Importance-ranked discovery
Gradient-boosted trees and mutual-information scoring surface the highest-signal transformations from thousands of candidates automatically.
Online and offline parity
A single feature definition compiles to both batch and streaming execution, eliminating train-serve skew by construction.
Governed lineage
Column-level lineage links each feature back to source tables, transformations, and the owners accountable for them.
02 · Probabilistic forecasting
Medovac Forecast
Demand, capacity, and revenue forecasts with calibrated uncertainty.
Medovac Forecast is a probabilistic forecasting service built for planning teams that need to reason about risk, not just averages. It automatically benchmarks a temporal fusion transformer against classical baselines such as exponential smoothing and ARIMA, then selects or ensembles per series based on backtested pinball loss. Every forecast ships as a calibrated predictive distribution with prediction intervals, so downstream planners can set safety stock, staffing, and budgets against explicit service levels rather than guesses.
Techniques
Key capabilities
Attention-based horizons
A temporal fusion transformer captures seasonality, holidays, and known future covariates across long horizons with interpretable attention weights.
Distributional output
Quantile regression produces calibrated prediction intervals, so planners can commit to explicit service levels.
Automatic baseline benchmarking
Every series is compared against ETS, ARIMA, and seasonal-naive baselines, and the platform ensembles only when it beats them on backtests.
Hierarchical reconciliation
Forecasts across SKU, region, and channel levels are reconciled so totals stay coherent from the leaf up.
03 · Anomaly and drift monitoring
Medovac Sentinel
Catch broken metrics, drifting models, and silent data outages first.
Medovac Sentinel is the observability layer for your data and models. It learns the normal behavior of every metric, table, and prediction stream using an ensemble of isolation forests and seasonal-hybrid decomposition, then flags deviations before they reach a dashboard or a customer. Sentinel distinguishes genuine incidents from expected seasonality, groups correlated alerts into a single root-cause narrative, and tracks feature and prediction drift with population stability and Kolmogorov-Smirnov tests so model decay never goes unnoticed.
Techniques
Key capabilities
Unsupervised detection
Isolation-forest ensembles and STL decomposition flag anomalies without hand-tuned rules or labeled incidents.
Drift surveillance
Population stability index and KS tests continuously compare live distributions against training baselines.
Correlated root cause
Alerts from related tables and metrics are grouped into a single incident with a ranked list of likely causes.
Adaptive thresholds
Bands widen and tighten with each metric's own seasonality, cutting false positives without missing real breaks.
04 · Causal inference and experimentation
Medovac Causal
Move past correlation and measure what actually drives outcomes.
Medovac Causal is a decision-science engine for teams that need to know why, not just what. It supports both designed experiments and observational studies, applying doubly robust estimation, propensity-score matching, and synthetic control methods to recover unbiased treatment effects. Causal surfaces heterogeneous effects with causal forests, so you can see which segments respond to a change and which do not, and it quantifies sensitivity to unmeasured confounding so stakeholders understand exactly how much to trust each estimate.
Techniques
Key capabilities
Doubly robust estimation
Combines outcome modeling and propensity weighting so an estimate stays valid if either model is correct.
Heterogeneous effects
Causal forests reveal which customer or patient segments respond most to an intervention.
Synthetic controls
Construct a data-driven counterfactual for launches and policy changes where randomization is not possible.
Sensitivity analysis
Quantifies how strong an unmeasured confounder would need to be to overturn a conclusion.
05 · Semantic analytics layer
Medovac Atlas
Ask questions in plain language, get governed, trustworthy answers.
Medovac Atlas is the conversational analytics layer that sits on top of everything else. It compiles a governed semantic model of your metrics, dimensions, and business logic, then uses a retrieval-augmented large language model to translate plain-language questions into validated queries against only certified definitions. Atlas never invents a metric: every answer is grounded in the semantic layer, cites its lineage, and returns the exact SQL it ran so analysts can audit and trust each result. Row-level security and column masking are enforced at query time, so self-service never becomes a governance risk.
Techniques
Key capabilities
Governed semantic model
Certified metrics and dimensions form a single source of truth that every question resolves against.
Grounded natural language
A retrieval-augmented LLM maps questions to certified definitions only, and returns the SQL it executed.
Lineage-backed answers
Every result cites the tables and transformations behind it, so trust is built into the response.
Query-time security
Row-level security and column masking are enforced on every query, regardless of who is asking.
See Medovac on your own data
Book a technical walkthrough with our field data science team. We will connect a sample of your data and show governed, production-grade intelligence in under an hour.