← All demos

NLP (on-device)

Zero-Shot Text Classifier

Route any text into labels you define, with no training, fully in your browser.

Production proof: The on-device version of the intent routing behind my production work (40+ keyword rules replaced by an LLM classifier across 22+ intents). Read the case study →

How it’s engineered
  • A ~26 MB MobileBERT model fine-tuned on MNLI runs on WebAssembly in your browser, scoring your text against labels you type in, with zero training or setup.
  • This is how modern intent routing works: define the intents in plain language, let the model classify, and keep it adaptable instead of hand-writing keyword lists that drift.
  • A model this small is excellent at distinct labels (sentiment, department, topic) but wobbles on lexically overlapping intents. Knowing that boundary is the job: in production I keep this tier for the easy 80% and route the ambiguous cases to a larger classifier or an LLM verifier, with a deterministic fallback for safety-critical routes.