Aquiles – PyTorch, LLM, Python

A comprehensive platform for identifying and optimizing medical journeys through various healthcare services. By mapping out treatments, appointments, and medical procedures, it enhanced patient navigation and ensures a seamless healthcare experience. The platform leveraged data-driven insights to improve care coordination, reduce inefficiencies, and provide better accessibility to medical services, ultimately enhancing patient outcomes. This was the third project Aquiles carried out for this company.

Aquiles was responsible for the following:

• normalized data from multiple sources (medical records, patient visit logs, billing records, etc.);
• implemented embedding models and Two-Tower neural networks to associate medical service descriptions with the most appropriate ICD-10 classification;
• tested Large Language Models (LLMs) to assess service coherence and determine if they belonged to the same medical case;
• leveraged tools like ChatGPT, Claude, and LangChain while studying advanced AI system design (e.g., Chip Huyen’s book on ML systems);
• created a system where multiple specialized agents work together - implementing LLMs with RAG, embedding models, and traditional ML classifiers that communicate through well-defined interfaces to deliver a cohesive healthcare intelligence platform;
• implemented model quantization (int8, int4), QLORA fine-tuning to reduce foundation model size while preserving performance, pruning techniques to remove unnecessary weights, tensor compression methods, and weight-sharing approaches;
• built and optimized custom transformer-based models using PyTorch and Hugging Face libraries.

Achievements:

• established a structured database containing essential information on each medical service;
• using semantic analysis, the system accurately identified disease codes based on medical service descriptions;
• the Two-Tower neural network successfully identified service similarities and grouped them into cohesive medical care lines.

The Data Request Bot was designed to streamline data-related inquiries by automating responses to frequently asked questions from the data team. It assisted with calculating specific metrics, performing clinical and statistical data analysis, generating reports and visualizations, and identifying relevant data sources. By improving efficiency and reducing manual workload, the bot enhanced data accessibility and decision-making processes.

Aquiles carried out this second project for the healthcare platform.

Among others, Aquiles' contributions included:

• carried out constant communication with stakeholders ensuring proper alignment of objectives;
• built a repository of business definitions and metadata was crucial to minimize hallucinations;
• built and optimized custom transformer-based models.

Achievements:

• the bot successfully answered basic analytical questions and generated simple visualizations;
• automated repetitive requests, allowing data scientists and analysts to focus on complex projects.

This was the third project Aquiles managed within cooperation with a healthcare platform: it aimed to develop a framework for identifying and reducing high-cost medical services to optimize the Medical Loss Ratio (MLR). The solution involved classifying medical services by category and subcategory, pinpointing the highest-cost sectors, and proposing cost-reduction strategies. These strategies included renegotiating supplier contracts, adjusting medical protocols, and optimizing resource allocation to improve financial efficiency without compromising quality of care.

Aquiles was in charge of the following:

• collaborated with insurance, actuarial, and finance teams was crucial for actionable insights;
• managed frequent adjustments and CFO engagement to ensure project alignment;
• developed expertise in MLR metrics and cost-cutting methodologies;
• developed a structured framework to segment medical services and facilitate cost analysis;
• created dashboards for executives and finance teams to focus efforts on high-impact areas.

The primary goal of this project was to design and deploy a recommendation system that integrated recipes with the product catalog available on Lider.cl. The chatbot, focused on enhancing customer experience, provided personalized recipes based on:

• Dietary preferences
• Purchase history
• Consumption trends
• Product availability in stores

• Increased sales conversion: The automated recipe-product pairing system significantly increased the purchase rate of suggested items.
• Enhanced user experience: The bot provided fast, accurate, and contextualized responses, improving customer satisfaction.
• Seamless integration with Lider.cl: Achieved real-time interoperability between the recommendation system and the e-commerce platform, ensuring up-to-date inventory and promotions;
• Built and optimized custom transformer-based models.

This initiative focused on defining and developing an end-to-end ML Framework embedding MLOps best practices for the entire model lifecycle, from exploration to deployment and monitoring.

• Heterogeneous tech stack across teams: Different teams used varied tools and formats, complicating integration.
• Scalability challenges: As the number of ML models grew, clear guidelines were required for versioning, deployment, and data governance.

• Established a standardized development workflow: Created a common repository, code review guidelines, unit tests, and CI/CD pipelines.
• Automated ML model lifecycle: Implemented pipelines for fast and secure deployment, reducing human errors and downtime.
• Developed internal documentation & best practices: Compiled a comprehensive guide on design patterns, infrastructure configuration, and scaling strategies.

This project aimed to segment and prioritize customers for email marketing campaigns using a purchase propensity model.

Key variables included:

• Purchase history and browsing behavior
• Previous email interactions (opens, clicks, unsubscribes)
• Sociodemographic data and website behavior

The main goal was to increase marketing ROI by:

• Focusing efforts on high-conversion customers
• Reducing irrelevant email blasts

• Enhanced campaign effectiveness: The propensity model significantly improved email open rates and click-through rates (CTR).
• Reduced unnecessary emails: Optimized marketing costs and brand perception by reducing low-relevance email sends.
• Seamless integration with CRM platforms: The model was automatically integrated into the email marketing system, enabling real-time segmentation.