I built a local AI coding assistant from scratch. No ChatGPT API. No cloud. Just raw engineering. cursor-lite runs entirely on local LLMs (Qwen2.5-Coder, LLaMA3.2) orchestrated through a custom Node.js agent I designed and built myself. What makes it interesting: → A multi-model orchestration layer that plans, validates, and executes code tasks step by step → A tool safety system that detects missing dependencies, asks for confirmation, installs them, and resumes, all without losing state → A ChatGPT-style streaming UI with live markdown rendering, syntax highlighted code blocks, and a typewriter reveal effect → Shipped using a production-grade Git workflow isolated branches, PR reviews, staged promotion with an AI agent doing the heavy lifting under my direction The part I'm most proud of? I used AI to build AI. I architected every system, wrote every prompt, reviewed every plan, and caught every bug the agent executed. That's a new kind of engineering workflow. And I think it's the future. Demo video below 👇 Drop a comment if you want to know more. #AI #LLM #SoftwareEngineering #LocalAI #BuildInPublic
The agent did the work, but you were the system that kept it coherent 💪. That’s the part most people skip when they say “I used AI to build AI.”
"I used AI to build AI" .. this is the part people underestimate. The value isn't in the code the model writes, it's in the system design, the guardrails, and the judgment calls the human makes. You're not replacing engineering, you're compressing it. Nice build.
Looks awesome, looking forward to seeing how this plays out...
A great step , this is paving way to intelligence democratization ! Keep Going
This is next-level! 🚀
Mayank Agrawal Strong build — the orchestration layer is where the real leverage is, and what stands out isn’t just “AI building AI” but you owning the planning, validation, and control loops, which is the real shift from user to system designer; would be interesting to see how you’re handling model routing, failure recovery, and evals, since that’s where these systems either scale or break.
Mayank Agrawal Stateful dependency handling + multi-model orchestration, fully local — that's production thinking. How's the latency holding up compared to cloud-based tools?
This is really cool, especially running locally 👏
The tool safety system with dependency detection and auto-resume is the part most people will underestimate. That's the difference between a demo and something you'd actually ship to production. We ran into the same challenge building our agent orchestration layer, the hard part isn't getting AI to write code, it's getting it to fail gracefully when the environment doesn't match its assumptions. Are you planning to open source the orchestration layer?