The Signal

A solo builder shipped AutoProber — an AI-driven hardware hacking arm built from duct tape, an old webcam, and a CNC machine. The system uses computer vision to autonomously probe physical circuit boards: locating test points, placing pro bes, and analyzing results without a human in the loop. 147 points on Hacker News front page. Zero VC money. Zero fancy lab. Just a GitHub repo and a will to ship.

This is the entire premise of indie hardware security research compressed into one project: physical + AI + automation = capability that used to require a team and a lab budget.

Builder's Take

Here's the leverage math that matters.

A professional hardware pent esting engagement runs $5K–$20K/ device. A hardware security engineer costs $150K+/ year. AutoProber collapses that stack to: one webcam (~ $30), a CNC frame (~$200 used), and API calls to a vision model (check current pricing for GPT-4o Vision or Claude 3.5 Sonnet).

The moat this creates is repeat ability at zero marginal cost. Once you've trained the system to recognize a PCB layout, it can probe 1,000 boards as easily as one. That's Naval's code leverage applied to physical labor.

The moat this destroys: bout ique hardware security firms charging for human-hours of probe placement. If you're a solo consultant in this space, AutoProber is your 10x productivity lever. If you are that boutique firm, you need to see this coming.

More broadly, this is a template for any vision-guided physical automation problem. PCB pro bing today. Automated product QA tomorrow. Pill sorting . PCB defect detection. Agricultural scanning. The underlying pattern — camera + vision LLM + CNC actuator — is domain-agnostic. One solo dev just proved the stack works with hardware from a garage sale.

The real insight: CNC machines are just programmable XYZ actuators. Pair that with a vision model that can read coordinates from an image, and you've built a general-purpose physical robot for under $300. The "robot " part was always the hard bit. Vision AI just made it trivially easy.

Tools & Stack

Core Hardware

  • CNC Frame: Any 3018 or similar desktop CNC (~$150–$250 used/Amazon). AutoProber uses it purely as an XYZ stage — no cutting required.
  • Webcam: Any USB webcam with decent resolution. The project literally uses an old cam. 1080p preferred for accuracy.
  • Probes: Standard pogo pins or grabber clips, ~$10–$20 from AliExpress.
Software & AI Stack
  • AutoProber: github.com/gainsec/autoprober — open source, Python-based. Clone and run.
  • Vision Model: GPT- 4o Vision or Claude 3.5 Sonnet for PCB analysis and coordinate extraction. Check current pricing at platform dash boards before budgeting.
  • CNC Control: GRBL (open source CNC firmware) via serial port. Standard USB-to-serial interface.
  • OpenCV: For camera feed processing. pip install opencv-python
  • pyserial: For GRBL serial communication. pip install pyserial

Quick Stack Check

# Clone the repo
git clone https://github. com/gainsec/autoprober
cd autoprober

# Install deps
pip install -r requirements.txt

# Connect  CNC via USB, configure port in config 
# Point webcam at target PCB
# Run
python autoprober.py

Alternative Vision Backends

  • GPT-4o Vision API : Strong spatial reasoning, good at identifying component labels. Check current pricing at platform. openai.com.
  • Claude 3.5 Sonnet: Comparable vision capability, sometimes better at structured output. Check current pricing at console.anthropic.com.
  • Local option: LLaVA or Moondream2 via Ollama if you want zero API cost and have a GPU. ollama pull mo ondream to get started.

Ship It This Week

Project : Automated PCB Test Point Logger for Your Own Hardware

If you build physical hardware products — even as a hobbyist — here's your weekend project:

Goal : Build a vision-guided probing rig that automatically maps and logs UART/JTAG/SWD test points on your own PCBs, and stores results in a local SQLite database.

  1. Day 1 — Hardware: Order or dig out a 3018 CNC. Mount a webcam to the spindle carriage with zip ties or duct tape (literally — see the source project). Connect CNC to laptop via USB.
  2. Day 2 — Software: Fork AutoProber. Swap the vision backend to your preferred API. Write a simple prompt: "Identify all exposed test pads on this PCB image. Return each as ( label, x_pixel, y_pixel)." Map pixel coords to CNC XY coordinates via calibration grid.
  3. Day 3 — Ship: Add SQLite logging. Run it against a spare board. Share results on X/GitHub. Tag it #buildinpublic.

Monetization angle: Wrap this as a S aaS for small hardware startups who need cheap pre -production board validation. Charge per board scan. Your COGS is API tokens + electricity. Everything else is margin .

The repo is live. The stack is proven . The only thing between you and shipping is a weekend.