Caio – Python, LLM, AWS

The client was a healthtech company focused on automating patient engagement and support. They developed a multi-agent workflow platform powered by large language models (LLMs) to streamline communication through voice and text, reducing manual workload and improving the efficiency of healthcare operations.

Caio was carrying out the following responsibilities:

• created a dynamic multi-step agentic framework with LangChain, enabling the chatbot to respond adaptively to complex user scenarios;
• integrated Langfuse for real-time logging and monitoring, which enhanced error detection capabilities by 40% and facilitated quicker model refinement cycles;
• established continuous feedback loops using Langfuse’s observability suite, providing granular insights into GPT-4’s decision pathways. Achieved a 20% improvement in LLM response accuracy through data-driven adjustments, demonstrating the value of real-time usage analytics in fine-tuning model performance;
• streamlined the end-to-end provisioning of LLM resources on Azure using Terraform, automating model deployments (e.g., GPT-4, Phi-3.5) and secret management via Azure Key Vault. Reduced setup time by 50%.

The client was a software consultancy supporting external clients with AI-driven solutions. They developed document parsing applications and contextualized chatbots powered by large language models (LLMs), enabling more efficient data extraction and conversational interfaces tailored to specific business needs.

Among others, Caio managed the following responsibilities: • spearheaded the creation of a fully automated deployment pipeline for an AWS Bedrock chatbot app & integrated unit tests specifically designed for AI, enhancing operational productivity and reducing deployment time by 70%;

• architected a document parsing application implementing AWS Bedrock and AWS Textract, automating the analysis of 200-page documents, which increased analyst productivity by 40% and eliminated 15 hours of manual review each week;

• led the design and execution of pre-launch user testing for an AI bot leveraging OpenAI's GPT-4 technology; collected qualitative insights from 40 participants that directly improved response precision by 25% and user retention rates;

• created a robust web scraping architecture with Python for gathering unstructured data across 50+ diverse domains; automated the integration process into MongoDB via Databricks, boosting overall performance and accessibility of collected insights;

• engineered and deployed AWS Lambda functions that utilized a FastAPI application to handle real-time data ingestion, integrating smoothly with PostgreSQL for instant storage; increased overall operational efficiency by 50%

The client was Brazil’s largest digital bank, focused on modernizing financial services through technology. The work involved developing ETL procedures and forecasting services to enhance data processing and support data-driven decision-making across financial operations.

Main responsibilities of Caio included, but were not limited to:

• developed an AI-powered forecasting tool using Python and stats models, now used by 1,000+ users, including C-level executives, to support strategic decisions;
• designed a data lifecycle protocol processing 500K+ points, improving insight quality and enabling faster analytics responses;
• built a scalable data ingestion pipeline with Clojure and AWS Lambda, handling over 1M records monthly and accelerating data access for analysts;
• optimized ETL workflows via Apache Airflow and custom Clojure/Spark Scala solutions, saving 15+ hours/month on routine tasks;
• scheduled and executed transformations in Databricks using Python and Spark SQL, producing reports for senior leadership.

Customer was a fraud prevention company that implemented biometric validation solutions to enhance the accuracy of identity verification and reduce fraudulent activity during digital transactions. Their approach leveraged advanced biometric technologies to ensure secure and seamless customer authentication.

Caio achieved the following:

• built an NLP app using TensorFlow to analyze user sentiment, uncovering product insights that boosted user engagement by 25% in six months;
• developed a facial recognition system processing 10K+ images daily, reducing fraud detection time by 60% and increasing trust;
• deployed a user monitoring system that informed key product tweaks, scaling usage from 500 to 2M+ monthly requests;
• designed ETL pipelines via Azure Data Factory and Spark Scala, improving data processing speeds by 30% in high-volume settings;
• created a scalable data architecture with MySQL and Azure Synapse, capturing 1B+ records/month and tripling query performance;
• rolled out a data governance framework establishing cost accountability across five departments, improving data accuracy by 40%.