About the role
We are looking for a student intern in Montreal to work on agent-driven 3D scene reconstruction. Given a posed image set (multi-view images with camera poses), LLM agents such as Claude and GPT models rebuild the scene as a clean, explorable 3D environment by calling 3D tools (Blender and the open-source Crucible engine) through MCP. These scenes serve as simulation environments for robotics.
This is a paid internship of about 2 months with a flexible start date. The role is hybrid, with 1–2 in-person days per week in Montreal. Fully remote is not available.
You will work directly with a researcher on a small, fast-moving project and own a well-scoped piece of it from design through evaluation.
What you will do
- Build and improve agent workflows that turn posed image sets into structured 3D scenes through MCP tool calls covering geometry, materials, lighting and rendering.
- Combine multimodal LLMs with real captures (images, camera poses, COLMAP output) so that agents rebuild real environments with high visual fidelity.
- Design tasks and metrics that compare models and toolchains on success rate, visual fidelity against the input views, number of tool calls and token cost.
- Find where agents fail, form hypotheses, run controlled experiments, and feed the findings back into prompts, tools and workflow design.
- Document results clearly and present them to the team.
What we are looking for
- Currently enrolled student (BSc, MSc or PhD) based in Montreal, in computer science, computer graphics, computer vision, robotics or a related field.
- Strong programming skills in Python or TypeScript, and the ability to drive a project independently, design experiments and analyze results.
- Project experience in at least one of: LLM agents and tool calling (including MCP), multimodal large models, 3D vision, computer graphics, or robot learning.
- Basic 3D understanding: cameras, coordinate transforms, meshes and materials.
- Interest in applying large models to real-world robotics problems.
Nice to have
- Hands-on experience using Claude, GPT or Gemini models for scene reconstruction, automated experiments or data generation.
- Research results, open-source projects or independent work you can show.
How to apply
- Email your CV and a GitHub or portfolio link to jizong.peng@dconstruct.ai with the subject "Internship Application – Name – School / Year". Add one or two sentences on the most interesting agent or 3D project you have built.