Jaime S.

• • Coordinated a production multi-agent AI system that translates natural language into SQL over advertising data warehouses, coordinating a pipeline of specialized LLM agents — each responsible for a distinct step: intent understanding, context retrieval, semantic layer generation, example selection, and SQL generation

• • Implement a semantic layer consistency engine that automatically detects and resolves duplicate or conflicting schema definitions produced by LLM agents across multiple data sources, ensuring generated SQL always references a single, coherent view of the data

• • Built a SQL validation pipeline: executes queries against the live warehouse, uses an LLM judge to verify results match user intent, runs business-logic rules, and applies chart-ready data transformations in a single automated flow

• • Designed configurable failure handling across the full query pipeline, allowing individual steps to fail gracefully with per-step policies that determine whether to surface a partial result or retry

• • Architected a serverless LLM dispatch system on AWS Lambda with a provider abstraction layer that decouples agent logic from any specific model — enabling seamless swapping between OpenAI, Gemini, and Anthropic without touching agent code

• • Built an ActionsAgent: a parallel LLM pipeline that evaluates ad entity signals like spend, revenue, roas, etc against configurable scenarios and generates fully-structured campaign recommendations — typed action parameters, signal-grounded rationales, impact estimates, and step-by-step execution plans — with a self-correcting generation loop that re-prompts on schema errors, followed by an LLM judge verifying action selection, signal grounding, and parameter rationale quality

• • Technologies: Python (asyncio, Pydantic, FastAPI), OpenAI/LiteLLM, AWS (ECS, Lambda, CDK), Docker, PostgreSQL

• • 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 workflows

- Integrated short-term and long-term memory systems, leveraging Redis and embedding-based retrieval.

- Designed and implemented Agentic NPCs with autonomy both within the game (actively making decisions to interact with the environment based on vision, memory, user interactions, etc.) and with real-world integration (e.g., Twitter, YouTube, Spotify) through MCPs and agentic workflows.

- Developed a Python service control layer to handle intelligent routing, error recovery, and service coordination between AI modules (e.g., voice, memory, orchestration).

Technologies: Python, Redis, Langfuse, ElevenLabs, FastAPI, Docker