One of my favourite tools. Knippex strippers. 12.62.180. Makes short work of prep'ing cables. Even older GPU cables that don't follow a standard in a specific Dell PowerEdge rack server.
I asked Claude Opus 5.5 to design a gear and set up the CAM for my LDO Milo 1.5 CNC milll using only open-source tools.
Opus 5.5 is more than capable, designed the gear, then the CAM and gave suggested feeds and speeds for my specific set up.
Note: My CNC is temporarily out of action currently, video shows other people's Milo LDO 1.5 making chips.
Claude Opus 5.5 just made a completely new PCB for PCBA with JLCPCB. STM microcontroller, usb-c, i2c and power.
This summarises the steps taken to make it.
Let me know if you want me to provide a longer video with more details.
Since I had existing CAD images to start with, I decided to test Claude Opus 5.5 against Grok 4.7 (xhigh) for animation creation. They started with the same images, hence they look similar but I do like the Opus 5.5 output a lot more.
Which do you think has more promise?
Love when more RAM arrives in the mail. My poor-man's DGX Spark build (a Dell Poweredge R730) now has 160Gb of RAM.
This will allow a Q6_0 quant on Qwen3.8-flash-next and a bump up in the Context Window. Should get 10-12 token/s.
A preliminary view of a second PCB done completely with Claude Opus 5.5. Only a first draft for review. This is to drive multiple ST3215 Digital Servo from i2c. A few things are missing but not bad Claude for first pass!
Claude Opus 5.5 seems to learn really quickly. Here is the second Kicad Schematic I asked it to build. I didn't need to repeat all the feedback from the first to get a useful artefact.
The more you use it for a task like circuit schematics the more it quickly understands.
Claude Opus 5.5 continues to amaze me. Claude Opus 5.5 created a 3D printable enclosure for an existing PCB. Then it created this promo video for it.
The PCB, the printable STLs and this animation are 100% just from prompting in Claude Code.
Interesting, every brand of TPU I have printed will absorb moisture whilst the print is happening, here is a demo of not-drying, some drying and a lot of drying for me:
Same filament roll, same printer, same settings
Letting Claude Opus 5.5 design actual PCB hardware. Prompted on Friday, sent to JLCPCB Friday evening and it is in Production Saturday morning (China time).
A project idea I never had the time to complete. Now done in a single afternoon.
This board allows hot-swapping batteries on a robot and measuring battery levels. Never have to stop ROS2 and robot development to recharge a battery! It is compatible with QWIIC and Piicodev.
My first Claude Opus 5.5 designed PCB by Claude Opus 5.5. This has been sent to JLCPCB for fabrication and assembly. Input: prompt >> Output: Assembled working board.
Can you guess what it does and what I am using it for?
This is an under-rated feature of using Claude Opus 5.5 to design a PCB. Here is a BOM that Opus 5.5 researched, checked stock levels, found the cheapest option and found replacement parts in stock if there was an issue. I didn't lose 2 hours of my life to component BOM admin!
I am working on an open source EE benchmark (TestEE Bench). In the coming days it should be ready for release. It runs on a text interface but just for fun, it visualises the real time testing in a web page:
The next version of llm-visuals has improved colour theme across all screens.
It also can show you the full logs across all sessions to show exactly how many tokens you have generated locally.
A little sneak peak of an open source Electrical Engineering Test Bench for Local and Cloud AI Models I am working on. It is called TestEE 😀 Benchmarks the agentic use of KiCAD, FreeCAD, NGSpice and Blender. End-to-end tests: component selection, schematic design, PCB layout, enclosure design.
Grok and Opus given the same task: design a USB temperature sensor PCB in KiCad.
Left: Grok 4.7 | Right: Claude Opus 5.5
Grok 4.7 went with a STM32 MCU + MCP9808, 52×24 mm PCB
Claude Opus 5.5 did a MCP2221A USB-I²C bridge + MCP9808, 30×26 mm
Which would you send to fab?
llm-visuals v0.9.0 is out: a live terminal dashboard for local LLMs.
New:
• Intel Arc GPU telemetry
• NVIDIA stats via in-process NVML, no nvidia-smi spawning
• Finds containerized vLLM/SGLang and relative GGUF paths
Thanks @Cyrag6@seth586!
Advice I keep giving friends before big conversations.
If you've got an interview, assessment, or cold call coming up, use Grok Voice to play the other side.
I set up a text file with notes, ask it to go easy, then hard, then give feedback. Sounds dumb until you do it once.
Practice still wins.
I have a Grok Bot searching for cheap RDIMM ram.
It tells me the market moved up... yep I bought all the RAM left at AUD$75 / 16Gb yesterday.
I was buying this for AUD$50 only 2 months ago.
Grok 4.7 is meant to be good at Electrical Engineering. So I gave it this small task in KiCAD. Just a USB-C temp sensor. The layout looks pretty good. The schematic was functional but not very human friendly.
Anyone have a special prompt for human friendly schematics with Grok?
I tested the range of local image generation with Qwen-Image 2.1 and found it handles every type of image I could think of, sci-fi, posters, photos, paintings, maps, technical drawings. Just check out these, all made by Qwen-Image 2.1
@QwenDevs have really given us another great tool, this time for making stunning images locally. It really seems to understanding lighting.
If you have a GPU you should try out some of your ideas for local image generation with it.
LLM-visuals v0.8.2 has just been released
New features thanks to three new contributors. Thanks @Cyrag6@melroy89 and @mrpliberty now we have AMD support, improved endpoint detection under Docker & WSL2, and improved support for LM Studio's llama-server.
Is it great that AI can speed up hardware development with nice little TUI based tools.
Gone are the days of printing debug statements to the terminal. Now a mini test suite.
Multiple local LLMs can be used together to develop a single program in Grok Build by asking Grok Build to orchestrate how each model is used.
Here is 2 servers, running 3 Qwen models over 4 GPUs make a single game:
llm-visuals just got a v0.8.1 release.
Now supports Windows, Mac (and still Linux) on x86 and arm. Binaries included.
Thanks to @sethx86 for naive vLLM support and @skalyanaraaman for improving model detection and @IronWolve for SLang support.
Qwen3.8-flash-next Q3_K_XL runs on average
12 token/s
with an Nvidia V100 16Gb VRAM + 128 Gb DDR4 RAM
This was my poor-man's DGX Spark build. Dell r730 with 128Gb of RAM + 16Gb VRAM and an SSD for engram.
Total cost is roughly $1000 to $1200 all up.