Privacy & architecture

Local-first by architecture, not posture.

Phaeth's privacy story is structural: the pipeline runs on your hardware, only encrypted structural metadata reaches the AI layer, and no LLM is ever in the inference path. Here's exactly what does and doesn't leave your environment.


Only metadata. Never rows.

When Phaeth's AI agent plans your pipeline, it needs to know what your data looks like, not what's in it. We send the structural skeleton (column names, types, statistical summaries) so the agent can reason. The rows themselves are processed on your machine and never sent anywhere.

📂Raw data
source
──▶
💻Local
processing
──▶
🔐Encrypted router
(metadata only)
──▶
🤖AI layer
(planning)
──▶
You confirm
before execution
──▶
🌐Trained twin
(local artefact)

// raw rows never leave your machine


How the architecture stays honest.

💻
LOCAL-FIRST

Runs on your hardware

Your laptop, your on-premises server, or your private cloud. Phaeth is a piece of software you run, not a SaaS that ingests your data.

🔐
ENCRYPTED

Metadata-only routing

The AI planning layer sees column names, data types, and statistical summaries. It does not see your rows or values. The router enforces this.

HUMAN-IN-THE-LOOP

You approve every load-bearing step

The agent proposes; you decide. Pipelines pause for confirmation at join plans, feature engineering, and target derivation. No silent transformations.

📦
REPRODUCIBLE

Hash-verifiable recipe

Every trained model is captured as a deterministic recipe. Same inputs, same hash, same predictions: replayable byte-for-byte, forever.


No LLM is ever in the inference path.

The AI agent plans the pipeline: it decides which columns to use, which model to pick, which features to engineer. Once the model is trained, that plan is captured deterministically and the model runs without any LLM in the loop. Predictions are produced by classical ML algorithms with hash-verifiable parameters.

This is what makes Phaeth's predictions reproducible, auditable, and safe to deploy in regulated environments. It is also why we can ship the recipe to a compliance team and say: replay it yourself, the bytes will match.


Three deployment targets. Same recipe.

The trained model is a portable artefact. You pick where it runs, and you can change your mind without retraining.

💻

Local machine

Stay on your laptop or workstation. Best for development, privacy-sensitive data, and small-to-medium datasets.

🏢

Your own server

Push the recipe to your infrastructure: on-premises, private cloud, or air-gapped. Best for team access and production workloads on regulated data.

☁️

Phaeth's managed cloud

We host the twin and run live ingestion for you. Best when you want always-on alerts and an API without operating the infrastructure yourself.


Built for regulated industries.

The combination of local-first execution, metadata-only routing, and hash-verifiable recipes makes Phaeth natively compatible with the requirements of SOC2, GDPR, and HIPAA-adjacent workflows. Self-hosted deployment is the recommended path for regulated environments: you control the data plane completely.

SOC2-READY GDPR-COMPATIBLE HIPAA-COMPATIBLE SELF-HOSTABLE AIR-GAPPABLE

For specific compliance attestations or to discuss a regulated deployment, reach out at [email protected].


Predict the future.
Without giving up your data.

Open the app and see for yourself: nothing in your environment leaves it.