aither

Python SDK for the aither platform - contextual intelligence and model observability.

Features

• Zero-code logging: Wrap any model and predictions are logged automatically
• Framework support: sklearn, pytorch, tensorflow, tinygrad, transformers
• Smart sampling: Captures sample + metadata from features (never blocks inference)
• Label correlation: Track ground truth with trace ID correlation
• Non-blocking: Background worker handles all API communication

Installation

pip install aither

Quick Start

import aither
from sklearn.ensemble import RandomForestClassifier

# Initialize
aither.init()  # Uses AITHER_API_KEY env var

# Train your model
model = RandomForestClassifier().fit(X_train, y_train)

# Wrap it - predictions are now logged automatically
tracked = aither.wrap(model, name="fraud_detector")

# Use normally
predictions = tracked.predict(X_test)

# Get trace_id for label correlation
trace_id = tracked.last_trace_id

Supported Frameworks

Framework	Example
sklearn	aither.wrap(RandomForestClassifier(), name="clf")
pytorch	aither.wrap(MyNet(), name="net")
tensorflow	aither.wrap(keras_model, name="tf_model")
tinygrad	aither.wrap(tinygrad_model, name="tiny")
transformers	aither.wrap(pipeline("sentiment"), name="sentiment")

Also works with sklearn-compatible libraries: xgboost, lightgbm, catboost.

Configuration

Environment Variables

export AITHER_API_KEY="aith_your_api_key"
export AITHER_BASE_URL="https://aither.computer"  # optional

Explicit Initialization

aither.init(
    api_key="aith_your_api_key",
    base_url="https://aither.computer",
    flush_interval=1.0,
    batch_size=100,
)

API Reference

aither.wrap(model, name, **options) - Recommended

Wrap a model for automatic prediction logging.

tracked = aither.wrap(
    model,                          # Any ML model
    name="fraud_detector",          # Required: model identifier
    version="1.2.3",                # Optional: model version
    environment="production",       # Optional: deployment env
    sample_rows=5,                  # Max rows to sample (default: 5)
    sample_columns=10,              # Max columns to sample (default: 10)
    features_fn=custom_extractor,   # Custom feature extraction
)

# Use the wrapped model normally
result = tracked.predict(X)
result = tracked(X)                 # Also works
probs = tracked.predict_proba(X)    # Also tracked

# Access trace_id for label correlation
trace_id = tracked.last_trace_id
all_trace_ids = tracked.trace_ids   # All predictions

# Access underlying model
tracked.model                       # Original model
tracked.classes_                    # Passthrough to model attributes

aither.log_prediction(...) - Manual Control

For custom pipelines or when you need full control:

trace_id = aither.log_prediction(
    model_name="my_pipeline",
    features={"amount": 150.0, "country": "US"},
    prediction=0.87,
    version="1.2.3",
    environment="production",
)

aither.log_label(trace_id, label)

Log ground truth for a prediction:

aither.log_label(trace_id=trace_id, label=1)

@aither.track(name) - Decorator

For functions instead of model objects:

@aither.track("my_function")
def predict(features):
    return model.predict(features)

result = predict(X)
trace_id = aither.last_trace_id()

Feature Extraction

Wrapped models automatically extract features with smart sampling:

# Input: pandas DataFrame with 50,000 rows, 100 columns
# What gets logged:
{
    "sample": [
        {"col1": 1.5, "col2": "A", ...},  # 5 rows
        ...
    ],
    "_meta": {
        "type": "pandas.DataFrame",
        "shape": [50000, 100],
        "columns": ["col1", "col2", ...],
        "dtypes": {"col1": "float64", ...},
        "truncated": True
    }
}

Supported types:

• numpy arrays
• pandas DataFrames/Series
• polars DataFrames/Series
• torch Tensors
• tensorflow Tensors
• dicts (common in transformers)
• lists

Custom extraction:

def my_extractor(X):
    return {"shape": X.shape, "mean": X.mean()}

tracked = aither.wrap(model, name="m", features_fn=my_extractor)

Usage Patterns

Basic Model Wrapping

import aither
from sklearn.ensemble import RandomForestClassifier

aither.init()

model = RandomForestClassifier().fit(X_train, y_train)
tracked = aither.wrap(model, name="churn_predictor", version="1.0")

predictions = tracked.predict(X_test)

PyTorch Model

import aither
import torch

aither.init()

class MyNet(torch.nn.Module):
    def forward(self, x):
        return self.layers(x)

model = MyNet()
tracked = aither.wrap(model, name="my_net")

# Both work:
output = tracked(input_tensor)
output = tracked.predict(input_tensor)

Ground Truth Correlation

# At prediction time
predictions = tracked.predict(X)
trace_ids = tracked.trace_ids  # List of all trace IDs

# Store trace_ids with your predictions
save_to_db(prediction_ids, trace_ids)

# Later, when ground truth is known
for pred_id, trace_id in load_from_db():
    actual = get_actual_outcome(pred_id)
    aither.log_label(trace_id, actual)

FastAPI Integration

import aither
from fastapi import FastAPI

aither.init()
model = aither.wrap(load_model(), name="api_model")
app = FastAPI()

@app.post("/predict")
async def predict(data: dict):
    prediction = model.predict(data)
    return {
        "prediction": prediction,
        "trace_id": model.last_trace_id
    }

@app.post("/label")
async def label(trace_id: str, actual: int):
    aither.log_label(trace_id, actual)
    return {"status": "ok"}

@app.on_event("shutdown")
async def shutdown():
    aither.close()

Management API

The SDK provides namespaces for managing your organization, API keys, and user account.

import aither

aither.init()

# Organization info
org = aither.org.get()
print(org.name, org.plan)

# Usage stats for current billing period
usage = aither.org.usage()
print(f"API calls: {usage.api_calls}")

# Current user
me = aither.user.me()
print(me.email)

# API key management (requires admin scope)
keys = aither.api_keys.list()
new_key = aither.api_keys.create(name="Production", scopes=["read", "write"])
aither.api_keys.revoke(key_id="...")

Why .get() instead of direct attributes?

You might wonder why aither.org.get().name instead of just aither.org.name. This is intentional:

1. Network calls are explicit - .get() makes it clear you're making an HTTP request. Hidden network calls on attribute access would be surprising and expensive.

2. Caching semantics are clear - The returned Organization is a point-in-time snapshot. You control when to refresh by calling .get() again.

3. Error handling is predictable - Exceptions from HTTP failures occur at the .get() call site, not on attribute access.

# Recommended: fetch once, use the snapshot
org = aither.org.get()
print(f"{org.name} on {org.plan} plan")

# Compare states over time
org_before = aither.org.get()
# ... make changes ...
org_after = aither.org.get()
if org_before.plan != org_after.plan:
    print("Plan changed!")

Data Format

The SDK uses OTLP (OpenTelemetry Protocol) to send predictions as spans with ml.* attributes:

Attribute	Description
ml.model.name	Model identifier
ml.model.version	Model version
ml.features	JSON-encoded feature sample + metadata
ml.prediction	JSON-encoded prediction
ml.label	Ground truth value
ml.environment	Deployment environment

License

MIT