Ines Montani 〰️
inesmontani.bsky.social
💥 Founder & CEO @explosion.ai 👩💻 Developing spacy.io + prodi.gy + ellf.ai 🐍 Python Software Foundation Fellow ⚡ Hacking @feministai.party 🧠 AI, Machine Learning & NLP 💼 linkedin.com/in/inesmontani 🐘 sigmoid.social/@ines 💥 explosion.ai 🌎 ines.io
👤 Joined on 10 October 2023
6720
Followers
70
Following
237
Total Posts
📜 Kullanıcı Geçmişi
- 10 Eki 2023, 17:20
🎉 Bluesky'e katıldı - İlk kullanıcı adı:
@inesmontani.bsky.social - 12 Kas 2023, 23:40
🔄 Profil bilgileri güncellendi
🚫 Blokladığı Kullanıcılar
❤️ Beğendiği Gönderiler
Your binary-decisions point is also how to scope the judge: once each step is a yes/no, score the LLM judge against your labels per decision, not overall. Where it agrees with you, let it own that call. Where it doesn't is where the few hundred examples should go.
Show Post
The failure I hit was the judge agreeing with the brief, and not hearing the music. An LLM scored a soundtrack I'd just rejected 9/10 twice, repeating the moods I'd asked for, and couldn't tell when I swapped a sad pad under a happy section. A small local audio model caught the swap.
Show Post
Agreed. The practical version we landed on: make the judge also say how close the call was. Calibrated by confidence band, high-confidence calls hold up, the near-tie band is weakest — so that's the one you route to a human.
Show Post
Self-hosted runners is such a clean mental model for this. Keeps the heavy data-private compute local while the app layer stays managed. Way less friction than asking people to run the whole stack themselves.
Show Post