Irati Hurtado

Meta AI

I worked on Meta AI by labeling and evaluating the AI model that powers the virtual assistant, and by comparing its performance to the performance of similar models developed by other tech companies. I also helped with Meta's internationalization efforts in order to launch a Spanish version of the product in various countries in Latin America as well as in the United States.

Prompt engineering for conversation design

This is a project I developed for a company that offers vehicle replacement services for insurers, brokers, and fleet providers. I used ChatGPT to create a testing dataset containing sample user utterances to evaluate the efficacy of an ML model that would classify users' intents into different categories.

Open D Express

Open D Express is a web-based chatbot that I developed for OpenDialog. Open D Express offers delivery services (e.g., scheduling the delivery of an online order). The chatbot was developed as a demo to showcase OpenDialog's software capabilities, and it is now part of OpenDialog Academy and OpenDialog's documentation.

Asa voice assistant

Asa is a voice assistant that helps patients schedule medical appointments via phone. This is a project I developed after taking the courses Conversation Design and Designing Conversations with Voiceflow and CDI, both by the Conversation Design Institute.

The BILinMID Corpus

The Bilinguals in the Midwest (BILinMID) Corpus documents the Spanish and English spoken in the Midwest of the United States by different types of bilinguals. Many of these speakers are heritage speakers, making this one of the few available heritage language corpora. The corpus is open-access and has an user interface to allow users to explore the corpus.

Research on structural priming

Structural priming is an interesting psycholinguistic phenomenon that consists of a tendency to repeat a linguistic structure one has previously been exposed to. Structural priming has been attested in human-human interaction, but also in human-computer interaction. I conducted some experimental studies to find out more about this phenomenon.

Haiku generator

A haiku is a type of short form poetry that originated in Japan and that consists of three phrases, where the first and third line have 5 syllables each and the second one has 7 syllables. This makes a total of 17 syllables per poem which follow a 5-7-5 pattern.

Using the natural language processing library spaCy and a dataset of random texts taken from Project Gutenberg, I developed a haiku generator in Python. The code extracts groups of words from the dataset and puts them together following the pattern for a haiku (5-7-5 syllables per line).

Automatic dialect classifier

Language detection algorithms help identify the language of a given text. For example, this type of algorithms could determine that a sentence like "Mary bakes a cake" is written in English, whereas "María cocina un pastel" is written in Spanish. This project goes a step further and tries to identify the particular dialect of a text. Specifically, by using machine learning, the algorithm can detect whether a Spanish sentence belongs to the Mexican or to the Peninsular variety of the language.

To train the model, text data from movie subtitles was used. In this case, all movies had been produced either in Mexico or in Spain. The machine learning classification algorithm used was a Naive Bayes classifier. All code was written in Python. You can also learn more about this project by reading a blog post I wrote about it.