Guadalupe Higuera

• Built and operated production platform powering agentic AI systems for 10,000+ users across 5 business use cases (policy, finance, investments) — agents leverage MCP server integrations, multi-step reasoning, single- and multi-agent orchestration, tool-calling, and database connectivity.
• Designed agent deployment and configuration infrastructure enabling non-technical business users to independently deploy and iterate on AI agents, reducing engineering bottlenecks and accelerating pilot-to-production timelines.
• Architected and own automated regression testing framework validating multi-step agent execution paths in a drag-and-drop AI agent builder during platform migration.
• Debugged and optimized AI agent testing pipelines, improving deployment reliability and reducing CI/CD time.

• Developed .NET backend services and React front-ends for internal sales tools and customer-facing applications; shipped production code across multiple releases.

iRacing AI Race Strategist

• Fine-tuned Llama 3.1 8B with QLoRA on 9,269 synthetic examples; eval framework on 55 cases showed score improvement 29.7 → 78.5 and hallucination elimination (48/55 → 0/55) via 12-metric weighted scorer and LLM-as-judge blind A/B comparison.
• Designed event-driven LLM orchestration with priority-based dispatch, per-event cooldowns, and state-machine transitions for race events — replaces naive per-tick inference. Generated training data using Claude API with category-balanced distribution across key racing scenarios.
• Engineered async telemetry → LLM → TTS pipeline (800–1,100ms latency) with GGUF q4_k_m quantization for local inference and deterministic fallback when LLM fails.

AWS Bedrock: Agentic Shopping Assistant

• Built autonomous shopping agent using AWS Bedrock Agent with custom action groups for natural language search and tool-calling.
• Provisioned full AWS infrastructure with Terraform: DynamoDB, Lambda, API Gateway, S3, IAM Roles.
• Developed FastAPI backend integrating Bedrock Agent runtime; Lambda functions query DynamoDB based on LLM-extracted parameters.

Titanic Historical RAG

• Built a retrieval-augmented search engine over the 1912 Titanic inquiry transcripts that surfaces contradictions between witnesses instead of hiding them.
• OpenAI text-embedding-3-large + Pinecone serverless for retrieval; Claude Haiku 4.5 with structured JSON output for pairwise contradiction verdicts with 0–1 confidence scores.
• Indexed over 3,300 pages of US Senate Inquiry and British transcripts across 68 witnesses (~40,000 chunks).

Relevant Coursework: Knowledge Representation & Reasoning, Artificial Intelligence, Bio-Inspired Computing