A 24.1M-parameter semantic model running inside your browser.

This portfolio is itself a small-model project. It downloads an open-source embedding model, runs inference locally with WebGPU when the browser supports it, and falls back to quantized WASM when it does not. Your question is embedded and compared with a curated bank of grounded answer cards; no answer-generation API or portfolio server receives the prompt.

• 24.1M-parameter Mixedbread embedding model optimized for retrieval.
• GPU acceleration happens on the visitor’s device through WebGPU—not on a remote GPU server.
• A Web Worker keeps model inference away from the interface thread.
• Semantic retrieval handles loose wording, misspellings, and intent instead of requiring exact questions.
• Authored answer cards keep biographical claims grounded; the model chooses relevant material rather than inventing facts.
• The same small-model mindset appears in SecureScribe’s separate 46M-parameter on-device clinical classifier.