project · 2025 · Data Scientist · SAP
Enterprise RAG for SAP Support
A production retrieval-augmented generation framework over the SAP Help Portal and enterprise case management, with a prompt framework powering case summarization, root-cause analysis, outbreak detection, and component prediction.
At SAP I design and run a production-grade retrieval-augmented generation framework for enterprise case management and the SAP Help Portal, built on LLMs (Llama, Mistral, and GPT-based models).
A prompt framework, not one-off prompts
I built a prompt framework that powers several SAP AI services from one place: ISM case summarization and root-cause analysis, outbreak detection for generating major-case narratives, and AI-based component prediction. A dynamic, structured prompt-template system handles the different document types the support world runs on (SAP Notes, KBAs, Community Articles) so the LLM reasons consistently across all of them.
Retrieval over rhetoric
Most of the leverage is in picking the right document, not writing fluent prose. I led an Auto Prompt Tuning (APT) initiative for the RAG pipeline: prompt-variation testing (grid and random search) plus reinforcement learning (PPO and bandit-based) to optimise the model’s document-selection accuracy.
Safe and explainable by construction
Because this runs inside an enterprise, the load-bearing work is in the guardrails: secure prompt engineering that segregates fixed from user-modifiable content to prevent prompt injection, reasoning chains that force the model to justify why it selected a document, and a modular service registry (Factory / Registry / Service Composition) to route and execute LLM workflows. Model lifecycle, versioning, and deployment run on Kubeflow and CI/CD with security (Mend) and code-quality (Sonar) gates.