AILobstersabout 2 hours ago
Lev Adds Escalation to Fast Classifier Models
16 min read
A developer built Lev, an offline decision engine exposing the same wire API as TypeSafe's Jev. It runs a ModernBERT encoder in about 95 milliseconds on a laptop CPU, then escalates low-confidence queries to a local Qwen3.5-4B model. Independent tests put Jev at 66 percent on a 150-passage benchmark, tying Claude Haiku 4.5.
Level
Hype check
Tap to vote and see what everyone thinks.