Guzman – Python, PyTorch, Scikit-learn

Geographic assignment modeling of genetic data for the conservation of jaguars and other carnivores.

• Designed pipelines for genotype data preprocessing.
• Trained a foundation language model adapted to genetic sequences (DNABERT-2) to encode feline genomic representation.
• Used transfer learning from the feline foundation model to train a geographic assignment model for endangered jaguars.

Forecasting models, a recommendation system, and advanced integrations of large language models (LLMs) for the client's platform.

• Contributed to the Likely-to-Sell model forecasting the probability of a property entering the market.
• Led development of a recommender system to power search, suggestions, and personalization across the platform.
• Designed and developed LLM production monitoring, performance evaluation, and continual improvement components.
• Worked on platform integration components (graph knowledge and memory) to improve LLM agent awareness and personalization.
• Built AB testing and evaluation systems to measure the impact of our product under hard-to-measure conditions.

A machine learning-based dynamic pricing system developed for a used electronics retailer in New York City.

• Researched and experimented on pricing modeling under challenging market constraints.
• Designed, developed, and deployed a dynamic pricing system from the ground up for a used electronics company.
• Achieved a 20% uplift in gross profit as measured by the causal impact method during the first 6 months of launch.

A large unstructured dataset representing grocery retail transactions. Within this project, it was standardized, and pricing and forecasting models were developed based on it.

• Established a product classification pipeline converting raw unstructured text into structured databases.
• Delivered foundational datasets improving the quality and reliability of client-facing products.
• Contributed to ad hoc pricing and demand forecasting projects.
• Researched, tested, and documented machine learning models and frameworks with a focus on NLP and forecasting.
• Manipulated big data using SQL and Apache Spark.
• Developed and fine-tuned classification models using Scikit-learn, Tensorflow, and PyTorch.

The project involved developing a machine learning-based system to forecast crime hotspots for the local police department. The objective was to divide the city into grids and predict the likelihood of each grid experiencing a peak in criminal activity on any given day.

• Utilized a specialized deep learning architecture known as conv-LSTM, which integrates convolution operations within the LSTM cell and is particularly effective for spatiotemporal modeling.
• Achieved a 50% recall rate with a precision of 25% of the model.