Howzer v2.2: English joins German as a first-class language
Not a translation layer. English messages are read by models calibrated for English, with their own thresholds and their own tone profiles, and every message is routed to the right one automatically.
By Howzer Team, Product · product ·
The lazy way to support a second language is to translate everything into the first one and carry on. It works until somebody writes that this was the last straw, which translates into German as a sentence about hay. So v2.2 does the other thing: English gets its own models, calibrated on English text.
- Native German language model
- Compound words handled as single ideas
- Formal and informal address distinguished
- Risk patterns learned from German text
- Model calibrated on English
- Idioms read as idioms
- Four tone profiles
- Risk patterns learned from English text
Nothing gets dropped
When a message arrives, its language is identified with a confidence score and it goes to the matching pipeline. If nothing scores high enough, which happens with very short messages and with mixed-language ones, it still gets a basic analysis using the closest match. No message is ever silently discarded for being hard to place.
Also in this release
0.800.85
Fewer suggestions, and the ones that appear are better.
- Root cause analysis makes noticeably fewer mistakes on ambiguous messages.
- Risk scoring was recalibrated against English phrasing rather than reusing the German thresholds.
- Two long-standing inconsistencies were fixed: risk scores that did not mean quite the same thing in each language, and confidence values that got clipped at the extremes.