Jaime – LLM, RAG, Python

An AI-powered Text-to-SQL engine that combines ad performance data and video creative insights to generate actionable recommendations and explain the reasoning behind them.

• Coordinated a production multi-agent AI system that translated natural language into SQL over advertising data warehouses;
• Implemented a semantic layer consistency engine that detected and resolved duplicate or conflicting schema definitions produced by LLM agents;
• Built a SQL validation pipeline that executed queries against the live warehouse, used an LLM judge to verify results against user intent, applied business-logic rules, and transformed data for charts;
• Implemented a chart generation pipeline using LLMs to display data;
• Designed configurable failure handling across the full query pipeline with per-step retry or partial-result policies;
• Architected a serverless LLM system on AWS Lambda to expose individual agents and support asynchronous processes;
• Built an Actions Agent that evaluated ad entity signals and generated structured campaign recommendations, action parameters, rationales, impact estimates, and execution plans with self-correcting validation loops.

A platform for creating, owning, and interacting with autonomous 3D AI agents that operate across games, virtual worlds, and other digital experiences.

• Built a Redis-backed internal memory system persisting tool calls, results, and arguments across workflow executions;
• Engineered stateful dynamic agentic flows with global request state, enabling LLMs to adapt tool usage without predefined paths;
• Improved initial streamed message latency from 5s to 2s via background task orchestration for parallel workflow.

A custom RAG-based solution designed to process insurance policy data across multiple countries and provide accurate answers to complex domain-specific queries.

• Designed and implemented a GraphRAG pipeline from scratch;
• Developed graph creation features, including entity extraction, relationship mapping, and chunk relationships;
• Built graph querying capabilities with entity-relationship similarity search and keyword-based retrieval;
• Implemented routing workflows using chain-of-thought reasoning for accurate response generation;
• Improved accuracy from 60% to 92–95% across multiple PDFs, enabling production deployment as an agent-assist solution;
• Leveraged PostgreSQL for similarity search, Python for implementation, NetworkX for graph logic, and FastAPI for backend development;
• Deployed the solution in a cloud environment using Web Apps, Cosmos DB, and Azure DevOps CI/CD pipelines;
• Migrated the solution to Azure, including graph creation, document extraction, prompt management, secret storage, and graph querying;
• Used Azure Document Intelligence to extract text from tables;
• Used Cosmos DB for prompt management and KeyVault for secret storage;
• Enabled Azure AI Search for graph creation, graph querying with NetworkX, and similarity search for relevant document extraction;
• Collaborated in an Agile SCRUM environment to ensure iterative and high-quality delivery.

An Agentic RAG solution for Python repositories that enhances code understanding and query answering through structure-aware retrieval and contextual analysis.

• Optimized chunking strategies for Python repositories by leveraging the language structure and schema;
• Defined function definitions as nodes and class definitions as parent nodes with methods as child nodes;
• Created additional relationships between chunks based on programming-language syntax and code references;
• Provided a richer context to the LLM model for answering code-related queries.

A Python library for structure-aware code chunking that leverages language syntax to optimize chunk sizes and preserve relationships between code elements.

• Built a Python library to create dynamic chunk sizes that kept complete functions, methods, classes, and code blocks together;
• Implemented chunk relationships beyond parent-child and previous-next.