In a consulting developer role for the startup Overlay AI, I applied the latest state-of-the-art in NLP to a number of domains, and worked to efficiently build an entire pipeline in the resource-constrained context of a browser extension.

I implemented the following layers to run entirely in a browser (low latency, zero external requests) on any text editing context:

• tokenization, POS tagging, and dependency parsing (using C++ to WASM compilation);
• lemmatization for WordNet alignment (rust-based);
• ontology (WordNet with external relations to Wikipedia etc);
• embedding spaces (custom, built from traversing WordNet graph and other sources);
• custom rule-based system for suggestions (written from scratch in rust, using rulesets from various sources and custom rules)

Demonstrations of live sentence dependency exploration:

Demonstration of live recommendations & ontologies

Further information about related work can be found under machine learning.