Playground
Play with the work, do not just read it
Three pieces pulled straight from my projects. Every one runs client-side, no backend. Each note says where it would slot into the rest of the site.
Phoenix 125M · Sweta multilingual
Tokenizer Playground real model
Type anything and watch it split into BPE tokens. Toggle the 32K English vocab against the 64K multilingual one and see Hindi and Kannada stop fragmenting. Loads my real Phoenix and Sweta tokenizers from HuggingFace and runs them in your browser.
↳ placement Lives best on the Phoenix / Sweta project pages.
Retrieval · neural embeddings, in-browser
Semantic Search real model
Search my projects by meaning, not keywords. Runs the bge-small-en-v1.5 embedding model and the cosine ranking entirely on your device, no server.
↳ placement Lives best on the Projects index as a way to filter the portfolio.
Evaluation · WinoGrande-style
Beat Phoenix 125M real model
Pick the better sentence ending, then see what Phoenix actually predicted. The 125M model itself runs in your browser (ONNX, int8) and scores each option by log-likelihood. It sits around 0.507 versus a 0.50 coin flip, and that honesty is the point.
↳ placement Lives best on the Phoenix project page next to the benchmark table.
Same engine, real pages
These drop into the existing design as-is
They already read the site palette and flip with dark mode. Next step is wiring each one into its home page above.