We build a system that answers user questions by searching across multiple documents. First, we embed all document chunks and compare them to the question’s embedding to find the most relevant pieces. Then, we prompt a language model with both the question and the retrieved chunks. The answer is generated with full respect to context, and returned to the user—efficient, scalable, and context-aware.

We propose a method to benchmark different Automated Speech Recognition models with our new collected testset. We also see that our new wav2vec2 model [utilizing language models & ngram] performs better on this dataset than other state-of-the-art models.

Utilizing the Granger Causality test, we understand how each telecommunication node [representing a province] influences and is influenced by other nodes over time. Data is of call, message and mobile data usage, reported daily for each province during a 5-year period. We infer both correlation and causality influence levels.

Using Machine Learning methodologies, we propose a web fingerprinting attack, in which we try to extract information about users' data, specifically which website the user is visiting. We then find the source of leakage using ML interpretation, and mask those parts. Afterwards, we try information extraction again. With each try, we see that information leakage reduces, and we learn about TLS's leaker parts.