Skip to content
← All projects
Academic

AgentForge

Local-first multi-agent platform for data analysis, code execution and text-to-SQL, orchestrated with LangGraph.

pandasweb-fundamentalsagentic-ailanggraphllm-inferencepythonfastapidockersqlrest

AgentForge is the platform behind my Computer Engineering bachelor’s thesis, developed under Prof. Roberto Navigli at Sapienza University of Rome together with Michele Pio Forlani and Salvatore Iavarone. It makes technical operations such as dataset analysis, Python execution and database querying available through a single natural-language web interface, with no programming or SQL knowledge required.

The system runs entirely locally and starts with a single Docker command. No data leaves the machine unless explicitly configured, and the LLM provider is swappable.

Highlights

  • Gateway: a FastAPI entry point that streams responses and intermediate events to the frontend over Server-Sent Events (SSE), and handles uploads, telemetry and a runtime Control Plane.
  • Supervisor: a LangGraph node that routes each request to the right worker in a routing phase, supports multi-step delegation across workers, and writes the final answer in a finalizing phase.
  • Workers: three specialised microservices that return structured summaries to the Supervisor.

Agents

  • Pandas Analysis: template-driven pipeline over CSV, Excel and JSON files. It classifies intent into 9 typed categories, validates column names with fuzzy matching, and never sends raw data to the LLM.
  • Code Interpreter: runs Python in an isolated Docker sandbox using an Actor-Critic loop, where a second agent reads the traceback and proposes fixes. Variable state persists across turns.
  • SQL Query: natural-language to SQL over PostgreSQL, Supabase, MySQL and SQLite, with human-in-the-loop confirmation. Destructive statements never run without explicit approval.

The frontend is built in vanilla JS with Web Components, with no framework dependency.