projects

CarmyGPT V2 (carmygpt.home.arpa)

• Based on Hermes Agent, allows for fully agentic workflows with deployable sub agents for different tasks and contexts

CarmyGPT agents breakdown

*Each of these agents have their own unique skills, memories, and personalities that distinguish them from each other. I initially had just one singular Carmy GPT agent, but I found that the number of skills and memories I was giving it was bloating its context a lot, making it unable to accomplish even simple tasks. * For each agent, I am using deepseek v4 flash, as it is the most efficient model for my use case, and I can run it through openrouter with ZDR hosting through deepinfra for a good mix between privacy, cost and performance

• **CarmyGPT (default) **
  • day-to-day chat interaction
  • basic tools like web search, web fetch, vision,
  • can call upon the other agents to assist with tasks
  • has a general knowledge of me and my life, but not of any specialized topics
  • thinking level - medium
• CarmyCode (coding agent)
  • an in-depth coding agent, equipped with a knowledge of my programming projects and the workflows I like
  • uses a mix of prebuilt skills and custom skills
  • thinking level - max
• CarmyNerd (research agent)
  • a specialized research agent I created to act as a replacement for notebook LM, whose capabilities I really enjoyed, but whose privacy posture I didn't
  • Uses a plethora of custom skills to integrate into my exact notion workflow

graph TD

Notes[Live Notes]

Notes --> ProjectA
Notes --> ProjectB
Notes --> ProjectC

ProjectA --> User-A[User Notes]
ProjectA --> AI-A[AI Outputs]
ProjectA --> Sources-A[Sources]

ProjectB --> User-B[User Notes]
ProjectB --> AI-B[AI Outputs]
ProjectB --> Sources-B[Sources]

ProjectC --> User-C[User Notes]
ProjectC --> AI-C[AI Outputs]
ProjectC --> Sources-C[Sources]
	```
- ***user notes***
	- *permissions: user read write, agent read*
	- *description: hand written notes by me that I take on the sources, and from the AI output*
- ***AI outputs***
	- *permissions: user read, agent read write*
	- *description: outputs are responses to user queries that synthesize information from the sources and user notes to provide useful ways to learn about the topic. information in outputs must be verifiable and have in-line citations from either the sources or user notes*
- ***sources***
	- *permissions: user read write, agent read write*
	- *description: contributed by me and the agent, can be PDFs, images, youtube transcripts, websites, etc…*
- *thinking level: max*

• CarmyOps (operations agent)
  • agent that has an understanding of my homelab, my home network, and other technical aspects of my life that are not strictly coding related
  • monitors the cron jobs on the homelab
  • thinking level: max

CarmyGPT V1 (deprecated)

• Lightly customized LLM intended to give my dog Carmy an english speaking voice
• Based on HuggingFace SmolLM2
• OpenWebUI for chat interface
• Custom system prompt, temperature
• Has access to carmyfacts.md for factual reference

models

cloud (via API)

• DeepSeek V4 Flash - Primary (Active)
• Kimi K2.5 - Vision (Active) I use OpenRouter to choose exactly which AI infrastructure I run on, and I specifically selected DeepInfra as my main provider as it offers ZDR support

local (deprecated)

• Ornith 9B (inactive)
  • The leading local small AI model for agentic development, does work much better than it’s competition but still falls short of my needs
• HuggingFace SmolLM2 (Inactive)
• Gemma 4 E2B (Inactive) I ultimately moved away from local models as I moved further into agentic development, it is very difficult to run simultaneous agents locally on limited hardware