Edge Data Agent (lfedgeai/EDA)

Data-on-prem, Code-on-the-Fly

Inputs: Data, User Intention, User Preference
Outputs: Applications / Services

Rule: The code in this repo should all be generated from data and prompts only (not including toolings and basic MCP services).
Data, user intention, and preference can be dynamic or static.

Goals

We focus on AI applications on the edge from both technological & business points of view.

We are monitoring and pushing the boundary of models' intelligence based on on-prem data and algorithms for accurate, useful, efficient, and proactive user experiences.

Background

Rather than requiring users to upload raw data to cloud platforms, end users should be able to define their tasks (clarify their intents and preferences) while keeping data locally and consuming services generated from their own data.

EDA converts a user’s own data into a local on-demand applications or agent service by leveraging the latest auto code generation capabilities of LLMs

• Agent code is generated on the fly the first time the user interacts with the database
• No coding knowledge (maybe even no installation needed if possible) is required for users to build and use
• In contrast to centralized AI platforms, users do not need to upload raw data
• The goal is to allow users to interact with, digest, or serve their own data on-demand without prior application development

Sandbox Data

• sandbox/data/ stores the data, agent_card, and input for evaluation
• Run local_evaluation.py to evaluate agent performance

About Tooling

(11/2025) Removed the tool folders but developers can use off-the-shelf code generation tools such as cursor, gemini cli, etc. (03/2025) we are leveraging the smolagents and llamaindex frameworks for reading/writing files from the local directory. The agent writes retrieval code automatically.About

Assumptions

1. The accuracy and robustness of LLM code generation improves to human level (SWEBench https://www.swebench.com/)
2. The cost of autonomous coding goes to negligible
3. The accuracy and user experience can be further improved by local/personal knowledge of the data base and the information of user query history

Branch Strategy

• Main Branch: Targets solving problems with prompts only, with minimum coding
• Working Branch: Targets practical solutions using current technologies (provides basic RAG code examples, code functions e.g., writing/reading files, etc.)