Juan Luis Ruiz - Tagle

While working in CARTO, I took full ownership of the ML models and optimization algorithms of the Local MX, a tool built for the Havas Media Group, which makes predictions of coverage and impacts at a very granular level and computes a selection of billboards that maximize such metrics.

The client's interest was to measure the impressions (number of visits) and coverage (number of distinct visitors) each of their billboards in Spain received weekly. They also wanted this information segmented by different categorical variables: type of day, hourly range, age, gender, and income level. For this, our models were trained on data from several sources (telco, SDK data, sociodemographic, POI, etc.). Then an optimization algorithm ordered the billboards best adapted to the target campaign.

I got involved in this project at a calibration stage, in which I:

• Tweaked the ML models and algorithms to align with client expectations
• Automatized background processes for telco data ingestion, automatic enablement of new billboards in the tool, etc
• Extended the usage of the tool within the Canary Islands by computing SDK routes on this region with OSRM
• Handled the communication with the client for all technical matters

During my time at CARTO, I had the privilege to work with one of the leading beverage companies in the world. They had a vast amount of sales data to analyze to calculate various KPIs related to their fleet, stock clients, and other business areas. At the time, processing this data and performing the calculations were done purely in Excel, which was time-consuming and prone to errors.

Together with my team, we launched a Spark cluster in Databricks to automate the KPI calculations. This allowed us to leverage the power of distributed computing and easily process the massive amounts of data the client was working with. I worked closely with their team to understand their specific requirements. Then I implemented the Spark-based solution that automated the calculations, eliminating the need for manual intervention and saving countless work hours.

TweetWars is a web app that enables users to compare two Twitter accounts by analyzing their latest 200 tweets and displaying insightful results in an interactive dashboard.

The tweets of both accounts are analyzed using NLP techniques, including sentiment and emotion prediction, topic modeling, and tweeting behavior statistics. These results are presented in a dashboard and sent to the paying user.

Despite its complexity, the system is fully autonomous and requires minimal maintenance on my part. It is comprised of multiple seamlessly integrated microservices which take care of payment processing, tweet fetching, sentiment inference, dashboard generation, email communication, and other tasks.

I did a thorough spatial analysis of the effects of the pandemic on retail stores during Black Friday in four different cities across the US using SafeGraph's human mobility data. I compared data for 2019, 2020, and 2021 to obtain insights into the evolution of footfall traffic and presented my results in a webinar and an article published on SafeGraph's blog.

I participated as a speaker in the Spatial Data Science Conference held in May 2022 at the Royal Geographic Society in London. This conference is among the most renowned congresses for geographic information systems and spatial data.

I presented the CARTO Analytics Toolbox, an SQL library for cloud data warehouses' spatial analysis and modeling.

A script that scrapes the BOE (the official gazette published daily by the Spanish government) in PDF format. It extracts relevant information about newly registered brands, including the registrant's name, telephone number, company website, etc. It also generates an excel file containing all the scraped data in a structured way.

I write about data science, geographic information systems, and math. Some articles have also been published in CARTO's blog as well as in Towards Data Science and Cantor's Paradise, two important publications on the Medium platform.

Some examples:
• Generating fake data with pandas, very quickly
https://towardsdatascience.com/generating-fake-data-with-pandas-very-quickly-b99467d4c618
• What to expect when throwing dice and adding them up
https://www.cantorsparadise.com/what-to-expect-when-throwing-dice-and-adding-them-up-5231f3831d7
• Scraping Google Search (without getting caught)
https://juanluisrto.medium.com/scraping-google-search-without-getting-caught-e43bb91b363e
• Can neural networks predict the stock market just by reading
newspapers?
https://quantdare.com/can-neural-networks-predict-the-stock-market-just-by-reading-newspapers/

I created a scraping master-slave system based on Google App Engine. The main problem of scraping is that sites can block your IP if they detect misleading behavior. As a solution, this system orchestrates from a local process a scraper deployed in Google App Engine. The main idea is to start scraping and redeploying the scraper to get a new IP whenever the current IP gets blocked.

I built a web scraper during college to help me study the Swedish language. Given a list of words in Spanish, it gathers their translations and example sentences. The software also generates exercises to practice.

I have experience in creating LLM-based chatbots for different use cases:

A couple of examples follow:

1. Youtuber Chatbot
I fine-tuned LLMs to mimic the style of a Youtuber (MKBHD)
For this, I scraped the transcripts of his YouTube videos and generated a dataset of conversations based on these transcripts.
I tried two approaches:
• Finetuning Llama2 7b as a quantized model and deploying it as an inference endpoint in Hugging Face
• Finetuning GPT3.5 from OpenAI via thier API.
I also gave the chatbot access to a vector database with info about the YouTuber's videos in order to make recommendations.

2. Doctor Chatbot
For a client, I developed a chatbot that would interact with a database of doctors in the US and answer questions about it.
It would transform the questions asked by the user into an SQL query and then return the results of the executed query to the user.

3. Firecitadel bot:
A Twitter bot that imagines stories and generates images using open-source models. (https://x.com/firecitadel)