This is a learning project for building mixed deterministic and agentic systems.

Pydantic

I'm using a lot of Pydantic tools for agent I/O, API request and response schemas, and pydantic-settings for configuration. Pydantic AI makes each LLM step return a validated model rather than free-form text. Retry functionality handles repeating failed steps or flows.

Agent Graph

Multi-step processing is modeled with pydantic-graph as a DAG. This is basically a flowchart to map out each agentic workflow with a corresponding visual indicator in the UI. Each node is a small, testable unit.

Python

FastAPI serves an async REST API plus WebSocket channel that streams live progress to the web client. The same underlying service layer is reused by a Typer CLI and a Textual TUI. The CLI tools make it easier to get fast feedback during testing.

Data

Persistence is SQLAlchemy to SQLite with Alembic managing migrations. AI Model calls can target Anthropic or OpenRouter, and I track per-run token and cost. I use fast/cheap models, find out if there's consensus, and then use more expensive models if there isn't agreement. So far this seems to work pretty well for my use cases.

Summary

The whole thing is a dashboard for monitoring a few multi-step agentic processes, and I try to keep as many deterministic tool calls in the loop as possible to reduce LLM hallucination.