6 min readSommay Khanna
Building LLM automation that does not quietly break
A model that is right 90% of the time is not 90% useful if you cannot tell which 10% went wrong.
- AI
- Automation
The failure mode of language-model automation is not a crash. It is a plausible, well-formatted, wrong answer that flows straight into the next step.
The fix is structural rather than clever prompting: constrain the output to a schema, validate it before anything acts on it, and give the model tools for anything it should not be inferring. If a value can be looked up, look it up.
Log every run. A workflow you cannot inspect is a workflow you cannot trust, and trust is the entire point of automating something you used to do by hand.
More writing
7 min read
Designing a rescue robot around constraints, not features
Power budget, radio range and payload weight decide the architecture. The feature list comes afterwards.
- Robotics
- Engineering
5 min read
What a line-following robot actually teaches you
The sensor array is the easy part. Everything interesting happens in the tuning — and in what breaks when the track surface changes.
- Robotics
- Competition