Operator Program›How it works

How Growt works

A technical overview for operators evaluating the platform — what data Growt accepts, what it measures, and what it tells you.

What types of data does Growt compare?

Growt works with any high-dimensional numerical data that can be represented as a matrix. Requests are capped at 200 MB with up to 10,000 samples and 4,096 dimensions per dataset — the practical limit depends on your embedding size:

Embedding model	Dimensions	Max samples per dataset
BERT, sentence-transformers	768	~8,000
OpenAI ada-002	1,536	~4,000
OpenAI text-3-large	3,072	~2,000
Large LLM / custom	4,096	~1,500
CLIP, ResNet	512	10,000 (cap)
Tabular / sensor data	< 256	10,000 (cap)

Need larger datasets? The audit endpoint accepts two datasets within the 200 MB combined budget. For very large corpora, contact us about file-upload ingestion. In practice this covers:

Embeddings

Output representations from ML models, encoders, and foundation models. The primary use case.

Tabular feature vectors

Sensor readings, lab measurements, financial indicators, clinical variables — any structured numerical data.

Time series

Represented as sliding-window feature matrices or pre-computed embeddings from a time-series encoder.

Spectral and geospatial data

Raw band values or encoded patches from satellite or airborne sensors.

Raw images and raw text need to be embedded first. Growt operates on the representation space — not pixels or tokens directly.

What does Growt actually measure?

Growt measures structural consistency between two datasets — whether the internal structure of one dataset is reflected in the other. This is different from both semantic comparison (which requires knowing what features mean) and behavioral comparison (which requires model outputs).

Given two datasets A and B, Growt analyses their internal structure and reports:

✓
Coverage gaps — Which regions of dataset A's structure dataset B never reaches — and vice versa. A model trained on A has never seen those regions.
✓
Dimensional shift — Which specific feature dimensions are driving the divergence between the two datasets, ranked by contribution.
✓
Per-sample anomaly scores — For every sample in dataset B, how far is it from anything in dataset A's structure. Identifies genuinely novel inputs.
✓
Recommendations — Prioritised, evidence-backed actions — not just a number that crossed a threshold.

Not semantic, not behavioral. Growt has no concept of what your features represent. It doesn't need model outputs. It works directly on the structure of whatever numerical representation you provide. This makes it applicable across domains — the same API call works whether you're comparing medical embeddings, satellite patches, or sensor readings.

Which industries and domains does it target?

The primary entry point is AI/ML validation — it's the easiest domain to articulate ROI in. But Growt's structural approach applies anywhere two data sources need to agree.

AI / ML pipelinesPre-deployment model validation. Does training data structurally match production data? Validate before you ship, not after it fails.

Healthcare and pharmaMulti-site trial equivalence and federated learning validation. Certify structural consistency across hospital sites before pooling cohorts.

Satellite and geospatialCross-sensor and cross-season alignment. Radar vs optical, multi-temporal, multi-mission — does the data structurally agree?

Manufacturing and QACross-line and cross-plant equipment drift. Are two production lines generating structurally equivalent output, or has one drifted?

FinanceRegime shift detection and model retraining triggers. Has the distributional structure of market data changed since the model was trained?

Robotics and autonomySim-to-real and sim-to-sim transfer. Does simulated training data structurally match real sensor readings? Does switching simulators preserve training distribution equivalence? Does domain randomisation actually cover real-world variation?

The operator program is specifically designed for domain experts who understand their vertical. We provide the structural analysis engine. You bring the domain knowledge, the customer relationships, and the brand.

Is Growt open-source or commercial?

Commercial SaaS with a white-label operator tier. The core method is proprietary — developed through years of research into structural data alignment. There is no plan to open-source the engine.

Clean API surface

Integration-friendly REST and MCP interfaces. The method is closed; the API is straightforward.

White-label ready

Custom domain, your brand, your MCP tool names. Your customers never see Growt.

Operator franchise model

You own the customer relationship and billing. We provide the platform and infrastructure.

Apply to become an operator Talk to sales