Nagen (@Nagen) header
Nagen (@Nagen) profile picture

Nagen

North Carolina, USA
Joined April 2009
Photo by Nagen (@Nagen)Photo by Nagen (@Nagen)Photo by Nagen (@Nagen)Photo by Nagen (@Nagen)Photo by Nagen (@Nagen)Photo by Nagen (@Nagen)Photo by Nagen (@Nagen)Photo by Nagen (@Nagen)Photo by Nagen (@Nagen)Photo by Nagen (@Nagen)
Pinned Tweet
I built my own Jarvis-style voice assistant. It's open source. 🎙️ "Hey Spark, rise and shine." → your calendar, inbox and GitHub, read out loud. Mac + Windows, MIT. #OpenSource #AI #VoiceAssistant
1
1
1
4
208
Replying to @TheEaglegon
We've conducted additional inference performance tests on the RK3588. Below are the results averaged across 8 test runs: Average Generation Speed: 15.81 tokens/s Average Time to First Text: 265 ms Model: Qwen2.5-1.5B-Instruct Quantization: W8A8 (INT8) Inference Engine: RKLLM, running on the RK3588's integrated NPU Input Prompt Length: 145 characters, approximately 31 tokens (including the text wrapper) A note on the measurement methodology: The 15.81 tokens/s figure is an approximate generation speed calculated by re-encoding the actual generated text using the official Qwen2.5 tokenizer and dividing the resulting token count by the generation time. It is not a direct reading of the NPU's internal decode statistics. We hope these additional measurements provide a more useful reference for evaluating the RK3588 in edge AI deployments.
1
20
"Hey Spark, summarize my open PRs." Same voice, three possible brains: Claude, Hermes Agent, or a local model via @ollama. Which would you pick for an assistant that can touch your tools: best answers, or nothing leaving the machine? #Ollama #LocalLLM #LocalFirst #OpenSource
8
Hey Spark, delete the old build folders.It names the MCP filesystem tool and the folders it will remove, then waits for a spoken yes. No yes, no deleteWould you let a voice assistant touch your disk without that pause?#MCP #AIAgents #VoiceAssistant
1
8
"Hey Spark, summarize my last 3 commits" Same voice, same wake word, and the brain is one line of config: Claude, Hermes Agent, or a local model via @ollama Which brain would you trust with a voice assistant that can touch your files? #Ollama #LocalLLM #LocalFirst
6
Hey Spark, summarize my last three commitsSame wake word, same voice, three possible brains: Claude, Hermes Agent, or a local model via @ollama. Swapping is one line of config.Which brain would you leave running all day?#Ollama #LocalLLM #LocalFirst #OpenSource
1
12
"Hey Spark, summarize my open PRs"The wake word runs on-device, the brain is yours: Claude, Hermes Agent, or a local model through @ollama - swap it without touching the restAnything that acts needs a spoken yes first#Ollama #LocalLLM #LocalFirst #OpenSource
1
11
Nagen retweeted
Physicist Nathaniel Craig shares how he worked with GPT-6 Astra to develop a candidate lattice definition of the Standard Model. It’s the first post on our new science blog: openai.com/index/step-toward…
594
373
287
3,821
864,958
Hey Spark, use your local brain. Same wake word, same tools — the model swaps to a Qwen on Ollama so nothing leaves the machine. Claude or Hermes Agent when I want the heavy agent tooling. Bring your own brain, no lock-in. @ollama#Ollama #LocalLLM #LocalFirst #OpenSource
1
13
Nagen retweeted
🔒 Check your repo's code scanning alerts without leaving the GitHub Copilot app. With the new GitHub security extension, you can ask Copilot to find, explain, and fix (or dismiss!) an alert. Now in public preview ⬇️ docs.github.com/copilot/how-…
93
32
50
482
100,214
Hey Spark, turn the thermostat to 68. It calls the Home Assistant MCP tool, says what it will change, waits for a spoken yes. Wake word + speech stay local on CPU; only the brain hits the cloud. Trust the spoken-confirmed version? @home_assistant#MCP #AIAgents #HomeAssistant
2
13
A 12-second walkthrough - wake word, spoken yes, Home Assistant acting on the command.
21
🧠🔥 18.3 tokens/s: Run DeepSeek-R1 locally on reComputer Industrial RK3588. What can the reComputer Industrial RK3588 do with a 6 TOPS #NPU? Here's a real-world demo: running #DeepSeek-R1-Distill-Qwen-1.5B entirely on-device with #RKLLM + #Docker. ⚡ ~18.3 tokens/s generation speed ⏱️ 0.57s time to first token 🧠 ~573 generated tokens ⌛ 31.88s total response time With a custom Web UI for real-time streaming chat and built-in performance monitoring, you can interact with the model while tracking inference speed, latency, and reasoning output. No cloud APIs. No remote inference. Just local LLM deployment on the edge. 🔗 Explore the demo and deployment guide: sensecraft.seeed.cc/ai-lab/e… 🛒 reComputer RK3576/RK3588: seeedstudio.com/reComputer-R… 🛒 reComputer RK3576 Module/Dev Kit/IO Board: seeedstudio.com/reComputer-R… #SeeedStudio #reComputerRK #Rockchip #reComputer #RK3588 #VisionAI #EdgeAI #AIoT #ComputerVision #reComputerAILab #AIDeployment #TheEasierPowerfulAIComputer #TheAIHardwarePartner
11
8
2
130
10,260
Hey Spark, what is the oldest open issue in hey-spark? GitHub answers over MCP, no dashboard, no clicking. Ask it to file one and it reads the plan back, then waits for my spoken yes. @ModelContextProtocol #MCP #AIAgents #OpenSource
2
10