README.md

DeepEngineer 🔬

A deep search agentic system for scientific and engineering workflows

This is a very early work in progress.

Objectives:

Webcrawler:

• functions for linkup and tavily
• function for wikipedia
• function for arxiv
• function for pubmed
• function for sciencedirect
• function for pdfs
• function for markdown

Agents:

• Agent that can work with markdown sources
• Agent that can search the web
• Agent that can analyse big pdfs and markdown sources
• Agent that can draw
• Agent that can draw in 3D
• Agent that can run mechanical simulations
• Scientific prompt for engineering tasks

Deploying:

• Building a front in React
• Connecting the front to supabase for authentification and monitoring credits
• Deploying the deepsearch agent on GCP Cloud Run.

And since I am paying, making sure that it is impossible to have an infinite bill.

🎯 About

DeepEngineer is an open-source project designed to build a specialized agentic system that assists engineers and scientists. The ultimate vision is to create an AI agent that can reason, plan, and execute complex tasks, effectively augmenting and accelerating the engineering and scientific workflow.

🚀 Vision

The project is inspired by the outstanding challenge in AI to address the complexities of numerical simulation-intensive science and engineering. As noted in recent research, this requires a sophisticated composition of scientific reasoning with the ability to operate specialized software and tools.

"Addressing the degree of complexity required in numerical simulation-intensive science and engineering workflows – which requires the composition of scientific reasoning with the ability to operate simulation software – remains an outstanding challenge." — FEABench: Evaluating Language Models on Multiphysics Reasoning Ability

This project aims to tackle this challenge by creating an agent with the following core capabilities:

⚙️ Current objective

It is likely that this project will not be better than the non specialized deepsearch solutions like OpenAI's or Magnus on scientific questions. Even though those solutions tend to focus on software development, they are generally very good at everything.

That being said, I hope to learn a lot and I see it as a personal challenge to deploy this solution and then iterate on it. Well see.

Getting starded

Installation

git clone https://github.com/your_username_/DeepEngineer.git
cd DeepEngineer
uv sync

🛠️ Tool Use

The minimum tools needed are:

• Web crawler agent
• Scientific paper analyser agent
• Drawing agent
• Coding Agent

Packages used

• smolagents
• crawl4ai

External APIs

• Mistral (OCR)
• Deepseek (LLM)
• Perplexity / Talily / Wikipedia / Arxiv