Implemented function and tool calling capabilities to enhance LLM-driven agents using Mistral and Qwen models. Integrated model-specific libraries and methodologies—such as Qwen Agent for Qwen models—to enable structured tool invocation, external data access, and dynamic decision-making within agent workflows. This work focused on building a more robust agentic RAG system, improving task orchestration, reliability, and contextual reasoning across complex retrieval and automation scenarios.

Designed and implemented a multi-agent hybrid RAG workflow in n8n to automate the creation of SAP Integration Suite iFlows. The system coordinated multiple AI agents that combined semantic retrieval from a Supabase vector database with structured reasoning via a Neo4j knowledge graph, enabling accurate and context-aware generation of integration artifacts. The workflow orchestrated agent collaboration, retrieval, and decision-making to produce complete iFlow structures, reducing manual effort and improving consistency across SAP integration scenarios.

Fine-tuned domain-specific LLMs for SAP Integration Suite use cases using quantization, parameter-efficient optimization, and Transformer-based training workflows. Applied model distillation techniques to generate high-quality thinking datasets, extracting chain-of-thought (CoT) reasoning behavior from stronger teacher models to improve downstream reasoning performance. Leveraged the Unsloth framework to accelerate training and significantly reduce compute requirements while maintaining model quality.

Conducted market and licensing analysis to select appropriate base and teacher models aligned with deployment constraints and commercial usage. Iteratively evaluated, optimized, and validated model performance to ensure improved accuracy, reliability, and reasoning consistency for SAP integration scenarios. Successfully packaged and published optimized models to Hugging Face for hosted inference and further deployment.