aayushkumar121

Interesting idea especially the “navigation vs selection” framing.

In practice, how are you measuring the +10pp gain? Are you using fixed eval sets or something more dynamic?

I’ve seen small models look better on benchmarks but regress pretty quickly once prompts/tools change slightly, so curious how stable these gains are over time.

That makes sense. Feels like the fallback path might end up being the default for a lot of real-world apps.

Have you thought about combining weak a11y signals + OCR to build more stable refs over time, or is that too brittle in practice?

Nice approach combining AT-SPI2 + grid fallback.

Have you run into issues where the a11y tree is incomplete (e.g. Electron apps)? Wondering how often the grid/OCR path becomes the primary path.

This feels like the start of a “people layer” for public companies—almost like a Bloomberg terminal but focused purely on executive movement.

Curious if you’ve looked at second-order signals yet (e.g. clusters of execs moving across the same companies)?

This looks pretty interesting. The idea of having a verification layer for agent-generated code makes sense as things get more automated. One thing I’ve noticed with LLM systems is that even when things pass tests, they can still behave unpredictably at runtime. Curious if OQP is mostly focused on pre-merge verification, or if you’re thinking about handling runtime behavior as well?

Interesting idea. I’ve mostly run into this kind of drift at the LLM output level rather than in the code itself even with similar setups, behavior can change quite a bit between runs. Curious if you’ve explored that side as well.