What FeatureGraph does
The bridge between your agents and your data.
FeatureGraph is set up with your team around your own science and engineering time-series data, and it turns that setup into a store that agents can query, instead of code that agents have to write.
Set up once, for your company. We work with your team to declare what the data contains and what the analysis should calculate and check. Those decisions are recorded in one place.
Analysis that's compiled, not vibe-coded. The compiler turns those declarations into explicit, measured objects. It is deterministic and runs outside any language model, so the same declaration gives the same result every time, and each result is recorded with the assumptions that produced it.
Queryable by agents and humans. A question in plain language becomes a short query that a separate program checks and answers from the stored results. Natural-language access is in development; the demo shows how it works on a small mock table. Changing an assumption, such as a smoothing window, creates a new defined analysis, visible and recorded instead of buried in a notebook.
FeatureGraph does not retain raw data in the service. Only the constructed summaries can be queried.
How it works
Analysis that's compiled and verified, not vibe-coded.
The first four layers are built. They are fully deterministic and run outside any language model. The fifth layer is in development for both human queries and agent-driven loops.
Primitive
A built-in compiler operation trusted by construction, such as identifying a rising or falling state, or calculating how much a quantity has exceeded a threshold over time.
Definition
A parameter or configuration chosen by the user or supplied from outside the compiler, such as the window size.
Object
The compiled output of a definition, such as the amplitudes, periods and durations of the individual waves in a sample.
Measure
A value computed from one or more objects, such as average duration or maximum temporal asymmetry.
Scientific interpretation in development
A claim about what the measures mean, written by a person or an agent. This layer sits outside the deterministic compiler.
The compiler is featuregraph-core, installed with pip install featuregraph-core. The current version is built for oscillating signals and can be adapted to other kinds of time series.
An example
Built-in representation of signal structure.
In an oscillatory signal, FeatureGraph reads a recording and splits it into individual waves. Each wave is called an object and is described by measures such as its height, duration and shape. In a respiration recording, each object is one breath.
The same construction applies to other oscillating signals, such as pressure changes in an industrial reactor or mechanical pulses from a rotating drill string. It depends only on the shape of the signal, not on outside context, so the method is not tied to one field. The accuracy of each measure traces back to the construction parameters, such as the smoothing window size and the detected period of the wave.
BIDMC and CapnoBase are two public respiration datasets. They measure breathing with different signals and sampling frequencies. FeatureGraph works in one-second intervals on both, so the same measures can be computed on each.