Ruan – LLM, Python, RAG

Three Azure-based commercial AI pilots for a large enterprise software client: automated document-type classification, OCR/LLM extraction, and stakeholder-facing review tooling. Covered a workflow processing 1M+ annual applications, from evaluation setup through production deployment and engineering handoff.

• Automated document-type screening with Azure Document Intelligence and LangChain, producing structured outputs with 95%+ accuracy.
• Course-corrected an underperforming OCR/LLM extraction pilot by building evaluation benchmarks and identifying Azure AI Foundry alternatives capable of 90%+ field-level accuracy.
• Built React-based reviewer tools, converted 40+ field corrections into prompt updates, deployed to Azure VM environments, and mentored 2 SWEs through 10+ knowledge-transfer sessions.

Product and technical delivery for a 6-month RAG proof of concept at a fintech platform processing 20% of all invoices issued in Brazil.

• Accelerated Weaviate vector indexing by 99%+ by replacing full-corpus retrieval with partitioned searches, discovered through BigQuery invoice data analysis.
• Optimized retrieval and ranking to 90%+ recall while cutting candidate depth from top-100 to top-20 — surfacing relevant matches with 80% fewer candidates.
• Interviewed 10+ candidates and mentored 1 SWE to enable the transition from successful PoC to production.

AI system suite for a Brazilian digital bank serving 30M+ customers, comprising a LangChain/LangGraph GenAI sales agent for card upgrades, a BigQuery-driven lead qualification pipeline, and a RAG-based bank product Q&A system — covering the full lifecycle from lead selection through conversational AI to product information retrieval.

• Developed a GenAI card-upgrade sales agent with LangChain and LangGraph, launching a 300+ user production pilot that cut customer acquisition cost by 6x while matching human-operator conversion rates.
• Used BigQuery analytics and time-series modeling to identify low-friction upgrade leads and feed qualified users into the AI workflow.
• Rescued an underperforming RAG system, beating OpenAI-based accuracy benchmarks and reducing inference costs by 90% through retrieval and prompt engineering improvements.

AI research and open-source NLP tooling across three organizations (one later acquired by Hugging Face), covering embedding-based annotation workflows, weak supervision, semantic similarity, legal entailment, sentiment analysis, and retrieval systems.

• Built Sentence Transformers + FAISS weak-labeling workflows with Snorkel, raising rule coverage from 32% to 79% and downstream model accuracy to 86%.
• Implemented embedding-based annotation features and contributed 68K+ lines of production-ready open-source code.
• Won 3 international NLP shared tasks in legal entailment, semantic similarity, and aspect-based sentiment analysis, beating runner-up teams by up to 6.4 F1 points.