Howzer v2.2: English pipeline goes live
The same analysis pipeline, now optimized for English-language customer feedback.
By Howzer Team, Product
What's new in v2.2
With v2.2, Howzer's pipeline supports English as a first-class language alongside the original German pipeline. This is not a translation layer. The English pipeline uses dedicated models calibrated for English text patterns, scoring thresholds, and response templates.
Key additions
- English-language pipeline with shared modular architecture.
- Multilingual fallback mode: unsupported languages get basic analysis instead of rejection.
- Language detection with confidence scoring and automatic routing to the right pipeline.
- Tone calibration profiles for positive, negative, neutral, and mixed messages.
Improvements
- Root cause analyzer precision improved, with fewer false positives on ambiguous messages.
- Risk scoring recalibrated for English text patterns.
- Response quality scoring threshold raised from 0.80 to 0.85.
- Deterministic response engine template coverage expanded.
Bug fixes
- Risk score normalization inconsistency between DE and EN pipelines.
- Sentiment model confidence output clipping at extreme values.
Language detection in practice
When a message enters the pipeline, the language detector scores confidence for each supported language. If confidence exceeds the threshold, the message is routed to the appropriate language-specific pipeline. If no language scores high enough, the fallback mode provides basic analysis using the closest match. This means the pipeline never silently drops a message. Every input gets processed.
- German BERT-based sentiment model
- Compound word handling
- Formal/informal register detection
- German-specific risk patterns
- English-calibrated sentiment model
- Idiomatic expression handling
- Tone profile calibration
- English-specific risk patterns