PowerPoint Slides Compliance Report Generation with AI
Role: AI Software Engineer
Company: Koltiva (Sugata)
Project Date: 2025
Project Overview
Developed an AI-driven application that automates the end-to-end generation of compliance report PowerPoint slides from custom data inputs. The solution applies robust software engineering practices and large language models (LLMs) to analyze user-provided data and transform it into structured, client-compliant presentations.
In addition, a human-in-the-loop (HITL) workflow was implemented, enabling compliance managers to refine and edit the generated PPTX through prompt-based requests. By replacing a fully manual reporting process, the application significantly improves productivity, reducing report creation time and operational overhead.
Features
- Custom compliance data is used to make the data.
- Agentic workflow that decides the layout, order of the slides, and which images to put together dynamically based on the data content
- Validation rules to consistently produce reliable slides. (e.g. check if there are content overflow)
These are some example slides during development. These are not part of the final product.
Key Contributions
- Developed a robust data pipeline using pandas, incorporating comprehensive data validation checks and automated artifact generation to support client expected PPTX slide creation.
- Designed and implemented an AI-driven pipeline leveraging LLM APIs and the Strands Agents Framework to automatically extract, analyze, and summarize compliance data for efficient PowerPoint slide generation.
- Ensured consistent and reliable PowerPoint slide output by utilizing structured data models with Pydantic and implementing validation scripts within the html2pptx workflow, integrating pptxgenjs, Playwright, and Sharp.
- Developed a Human-in-the-Loop (HITL) workflow, enabling compliance managers to conveniently edit PPTX files through prompt-based interactions when manual adjustments are required.
- Conducted comprehensive evaluation and observability of LLM outputs using LLM-as-a-judge methodology and integrated monitoring with Langfuse.