Hey everyone 👋 I think in Hinglish. My job doesn't. There's a real tax in translating "yaar this project is a mess" into something that survives an HR audit, and I was doing that translation manually, in my head, every time before hitting send on anything honest. So I built this to do it for me. The extension itself is simple - click, get a diplomatic rewrite, review the diff, decide. What's underneath it is the part I actually care about. It's not a GPT wrapper with a "be professional" system prompt. I'm using activation steering, you go into the model's hidden states mid-inference and physically push them toward a "diplomatic" direction, permanently, via forward hooks on specific layers. No fine-tuning, no retraining, no clever prompting. This took a few failed attempts to get right. Single-layer steering gave me either nothing (barely different from my input) or garbage (the model basically had a stroke). Multi-layer steering across layers 19-22 was what finally got me something that shifts register while staying coherent. And the vectors themselves are only as good as the contrastive pairs behind them, a lazy, noisy set produces sentences that technically sound "corporate" but mean nothing. It's live, free. If you also think in Hindi/Hinglish and rewrite your brain in real time before every work message, try it and tell me where it breaks.

Being Corporate

Because you can't always be human at work.

Because you can't always be human at work.