SMALL-MODEL ENGINEERING
The portfolio answers you locally.
No prompt is sent to an answer-generation server. A compact semantic model runs inside your browser, understands the intent of a question, and retrieves the most relevant grounded material about Anson’s work.
What happens after you press Send
Interpret
Your wording is normalized so loose phrasing and common misspellings still work.
Embed
The 24.1M-parameter model converts the question into a semantic vector on your device.
Retrieve
That vector is compared with grounded cards covering Anson’s projects, experience, and skills.
Answer
The closest supported response is assembled and streamed into the conversation.
Local does not mean simplistic
The system does not depend on an exact phrase lookup. It combines semantic similarity, lexical signals, project-aware routing, recent conversational context, authored intent examples, and supporting-card selection. That lets “secure scrap,” “clinical documentation thing,” and “privacy tool for nurses” converge on SecureScribe without pretending the model knows facts outside the public answer bank.
GPU means your GPU.
When WebGPU is available, inference runs on the visitor’s graphics hardware. Otherwise, the site falls back to an 8-bit WASM model on the CPU. Cloudflare serves the static files, but it does not generate the portfolio answers.
A pattern Anson uses beyond this website
SecureScribe applies the same philosophy in a higher-stakes setting. Its separate 46M-parameter behavioral-health classifier is designed for on-device clinical inference, while deterministic software constructs documentation and nurses retain approval. In both systems, the goal is not to imitate a giant general-purpose chatbot. It is to build a smaller, bounded model that performs a specific job privately, efficiently, and measurably.
Why this is a portfolio project
- Browser inference and Web Worker orchestration
- WebGPU acceleration with resilient WASM fallback
- Quantized model delivery and client-side caching
- Curated knowledge architecture and intent design
- Embedding evaluation, thresholds, and retrieval benchmarking
- Privacy-by-architecture: prompts remain on the visitor’s device
Measured, not guessed
The retrieval system is tested against a 200+ case benchmark dominated by short prompts, clipped wording, aliases, misspellings, ambiguous follow-ups, and unrelated negative controls. Separate tuning, holdout, and successive shadow sets expose overfitting. The optimization command embeds every test question, grid-searches scoring weights and the confidence threshold, records every miss and false positive, applies the winning configuration, and runs regression tests before deployment.