I am experimenting with EdgeAI Forge, an early-stage local-first architecture for using specialized AI agents in industrial vision, ROS 2, robotics, and edge-deployment workflows.
The intended split is:
- local models for frequent coding, planning, documentation, private-data processing, and visual reasoning;
- optional cloud escalation for difficult or large-context tasks; and
- explicit human review before any physical deployment or machine-control action.
The current proof of concept routes Planner, Vision, and ROS prompts to Ollama. I am looking for practical model recommendations and evaluation ideas for:
- local coding agents that produce structured Python/C++/ROS 2 artifacts;
- VLMs for industrial image inspection and pipeline selection;
- smaller reasoning models for planning and validation;
- embedding models for project memory and retrieval; and
- runtimes that work well across workstation GPUs and Jetson-class devices.
I am especially interested in measured tradeoffs—VRAM, latency, structured-output reliability, quantization, tool use, and license constraints—rather than leaderboard scores alone.
Project: GitHub - afridali123/EdgeAI_Forge · GitHub
Longer background: EdgeAI Forge: My Journey Toward a Local Agentic AI Platform for Industrial Automation
This is an early architecture discussion, not a finished product announcement. Suggestions for models, datasets, evaluation harnesses, or related projects are welcome.