Less Clueless
Engineering, products, and the thinking behind building things well.
I'm Ankur Goel. I build, ship, and write down what I learn along the way. Some of it tactical, some of it personal experiments.
-
DROID audit: Dataset's defect rate depends on which LLM you ask
Three models judged the same 72,851 DROID annotations with one rubric. The defect rate swung 2.9× — why a single-judge number measures the judge, not the data.
-
How to ship a “hobbyist” robotics kit?
Four gotchas from assembling an SO-101 kit, and what shipping good open-source hardware actually requires from vendors and buyers.
-
Pi0.7: utilizing bad data to teach good things
How π0.7 turns failures and low-quality demos into training signal by labelling them honestly, and whether that's a lasting paradigm or a stopgap solution.
-
When AI hallucinates what it sees: notes from a robotics sim
Three times Claude was confidently wrong about what it was seeing in a robotics sim, and what that means for AI moving into the physical world.
-
Building a personal automation system on the graveyard of past attempts
Building a personal automation system: a journaling and cataloguing loop on Notion, Telegram, and Claude.
-
Design, Launches and Doing it right
Lessons from two years building Chaotix: why design is trust, why shipping isn't launching, and a tactical framework for small teams.