DevObsessed

A studio can’t clone a voice actor’s voice without it in the contract. So can your company capture every judgment you make and turn it into an agent? Best debate from the AI Engineer Omaha Q&A. Where do you land?

“Your Claude is better than mine.” Same model. Different context. Brad Risse on wiring Claude into your notes, transcripts and the people you learn from, and why you shouldn’t keep that to yourself.

Your team makes hundreds of judgment calls a week. Where do they go? Brad Risse: save what the agent suggested and what the human changed. That’s your eval set. Stop tuning prompts. Start capturing judgment.

Hot take from Brad Risse: the harnesses you built for older models are now slowing the new ones down. He threw his out. Vaguer instructions, less babysitting, more steering up front. When did you last rethink yours?

Tom planned a “not hot dog” demo. Then found out the model was text only. The day before his talk: “just kidding, it was multimodal all along.” This is how fast things move right now.

Live PII redaction on transcripts. No training. It caught Persian names and international addresses. And it knows “1600 Pennsylvania Avenue” on its own isn’t personal data, but your home address is. Tom McLaughlin at AI Engineer Omaha.

You can’t write a regex for every command your agent might run. Tom McLaughlin’s idea: ask a fast classifier in plain English. Is this safe? Will it leak secrets? Can we undo it? Cheap enough to run on every call.

Ever asked an LLM how confident it is? It reads its own reasoning back and says “pretty confident.” Every time. Tom McLaughlin on why Jev’s confidence score is just math, not vibes.

We used to fine-tune BERT to classify and route text. Tom McLaughlin tried Jev from TypeSafe AI. Zero fine-tuning. Beat our old setup. Fast and cheap. From the first AI Engineer Omaha meetup.

Meet our very own Kalvin Tuel: youtu.be/lyaU2J6gIco?feature…