AIsra

AIsra – Agent for Interpretable Symbolic Regression and Analysis

What is AIsra?

Warning

This is not a production-ready application!

It is a prototype meant to demonstrate possible applications of LLM-powered AI agents in the area of prescriptive analytics and symbolic regression.

AIsra is an AI agent that can...

• train symbolic regression models on real-time data streams.
• react to concept drifts in the data stream by retraining the model when necessary. This is done by first training a single model that never changes ("base model"), and a second model that is retrained on the residuals of the base model when a drift is detected ("residual model").
• store the trained models in a Redis database and reuse them when possible to avoid unnecessary retraining.
• evaluate metrics of an SR model—such as model quality and feature importance—and use them to make decisions about how to improve the model and provide information to the user.
• explain certain aspects of the model as well as its decisions to the user in natural language.
• respond to user queries in natural language.

Furthermore, there is a frontend for interacting with the agent and visualizing the data stream, the trained models, and the metrics of the models in real time.

Running AIsra

With Docker

Prerequisites

• Docker
• .NET SDK 10

Steps

1. Copy the .env.example file in the root directory to .env and update the environment variables as needed (LLM_API_KEY and, unless disabled, the authentication settings are required).
2. Run run-docker.sh (on Linux/macOS) or run-docker.ps1 (on Windows) to build and run the Docker containers. Running docker compose up directly also works; the cert-init service generates the development certificate if it is missing.
3. Access the application at the specified port in your web browser (http://localhost:3000 by default).
4. Supply the agent with a data stream over MQTT (e.g. with janzenisek/dsg).

Authentication

By default, users have to sign in with a Google account. To set this up, create OAuth credentials in the Google Cloud Console:

1. Create (or select) a project and configure the consent screen under Google Auth Platform → Branding. Choose the External audience and publish the app under Audience; while it is in Testing, only the listed test users can sign in.
2. Under Clients → Create client, choose Web application and add the authorized redirect URI ${PUBLIC_API_URL}/api/signin-oidc (e.g. https://localhost:5003/api/signin-oidc).
3. Set GOOGLE_CLIENT_ID, GOOGLE_CLIENT_SECRET, and FRONTEND_URL in your .env file.
4. Set AUTHORIZED_USERS_REGEX to a RegEx matching the email addresses of users who should get access (e.g. ^[^@]+@example\.com$). Only email addresses verified by Google are considered.

Users without a Gmail or Google Workspace address can create a Google account for their existing email address at accounts.google.com/signup (Use my current email address instead).

Set USE_AUTHENTICATION=false to disable authentication.

Troubleshooting

If a container exits with Access to the path '/certs/localhost.pfx' is denied, the host path certs/localhost.pfx was created as a directory (usually because docker compose up ran before the certificate existed). Remove it and start the containers again:

rm -rf certs/localhost.pfx
docker compose up -d

The cert-init service regenerates the certificate on startup.

For Development

Prerequisites

• Docker
• .NET SDK 10
• Node.js 24
• pnpm

Steps

1. Copy the .env.example file in the root directory to .env and update the environment variables as needed (LLM_API_KEY and, unless disabled, the authentication settings are required).
2. Run the required services.

docker compose -f compose.dev.yml up -d

3. If you use the Grpc client type (default), run the AIsra.HeuristicLibWeb.Server project.

dotnet run --project ./backend/src/AIsra.HeuristicLibWeb.Server/

4. Run the web server project.

dotnet run --project ./backend/src/AIsra.Web/

5. Run the SvelteKit development server.

cd ./frontend/
pnpm install
pnpm dev

6. Access the application at the specified port in your web browser (http://localhost:5173 by default).
7. Supply the agent with a data stream over MQTT (e.g. with janzenisek/dsg).

Documentation

Click here to view the documentation.