Devin Willis

I’m rebuilding this account around what I actually do now: I build AI + robotics systems for the physical world. PCB design. Embedded C. Robotics control. ML/computer vision. Mobile apps. Medical-device engineering. Mechanical prototypes. The messy stuff where software has to survive contact with hardware. I’ve founded and built projects across pathology imaging, automated agriculture, medical devices, AI training tools, PCB automation, and robotics. I’m currently building PromptPCB and leading engineering on AutoTQ. This account is going to be build logs, technical notes, failures, fixes, prototypes, AI hardware workflows, and lessons learned the hard way. No guru content. No fake founder theater. Just engineering, product building, and the grind of making difficult systems work.

@seveibar Hows my autorouter attempt - Signals: 835 / 1,077 - Plane stitches: 731 / 731 - Unrouted: 242

trying to humble the autorouting team, TBD if it'll work

Have something big in the works...

so my usb super speed cable is breaking my GPS readings thats fun.

Almost time for plywood

@realwillreil This thing is so nice

I need to design enough low power PCBs to justify buying one of these bad boys

Work in progress

First results of ppo swarm based placement Not clear if it's actually doing something intelligent but probably not yet it's been under minimal training

Hello

We have a plasma cutter yet I'm still out here manually prepping ibeams

Video is a fundamentally different medium than text and it's not until the last year or two that the embedding space of these two were unified As powerful as modern LLMs are a model trained specifically in the embedding space focused on circuit design may actually make better use of the power of transformers

Follow up thought Embeddings models are the heart of what makes LLMs powerful, encoding what makes the rich semantic meaning of words into numbers. Creating an embedding model that takes in netlist information and produces embeddings that can be used to advise a transformer based placer or RL model could also be powerful It could potentially condense footprint information, electrical requirements, functional groups, component dependencies, etc into something that could actually be trained and differentiable rather than generic json

@raulizahi I am now

@DevWillTech Not a bad idea. Working on that?

Why can't every PCB component self assemble on a PCB layout like an RL agent swarm... I'll be right back

@ymuchosmemes must film on set

"El hombre nunca fue a la Luna, todo eran gráficos por computadora" Los gráficos por computadoras en 1970:

@seveibar Chefs kiss

@DevWillTech Could use some documentation but we have put a lot of cycles into this [1] (MIT open source)! Ofc you will need to convert kicad to circuit json i think [2] [1] github.com/tscircuit/copp… [2] github.com/tscircuit/kica…

Algorithm pours are gonna make me go insane

@seveibar Got someone on it...

@embedrapp This was gold😭 RTK enabled too obviously

@DevWillTech jetson + realsense + zed f9p? that robot is about to have better spatial awareness than i do.

New computer unit for the robot😁

DIY or Buy the LumenPNP v4? Anyone know the right answer?