Skip to content

PPT-Werkstatt

PPT-Werkstatt generates PowerPoint slide decks from uploaded source documents. An agentic pipeline composed of several specialised LLM agents segments the material, plans the slide structure, assigns layouts, fills the slides with content, and validates the results in multiple stages for material fidelity. The finished presentation can then be revised step by step through chat.

At a glance

  • Produce a coherent slide deck from a set of documents — the material is segmented and converted into a presentation-ready form.
  • Adapt the slide deck to a specific audience, language, and presentation duration — slide count and content depth follow the briefing.
  • Provide a rough outline or thematic priorities and have them reflected in the resulting deck.
  • Revise slides through chat — shorten, rewrite, reorder, add, replace, and undo as needed.
  • Have icons and diagram graphics added automatically — chosen from a curated set to match the slide content.
  • Use a custom PPTX design — master layouts, fonts, and colours from an uploaded template are preserved.
  • Export the deck as PowerPoint, Word, or Markdown.

Highlights

In contrast to a single LLM prompt, the application uses a multi-stage pipeline of specialised agents. This improves consistency, source attribution, and traceability of the resulting slides compared with a single model call.

  • Agentic pipeline — A five-phase process (structure, layout, content, quality, homogenization) distributes the work across independent agents. Each agent has a narrow scope of responsibility; intermediate results flow through the pipeline along a defined data path.
  • Multi-stage quality assurance — Generated slides are checked in three stages: rule-based (slot constraints), embedding-based (material coverage), and LLM-based (hallucination check). Slides flagged as problematic enter an automatic correction loop.
  • Two coordinated LLM roles — A primary model with thinking enabled handles quality-critical tasks (routing, structure planning, layout selection, validation); a faster secondary model handles parallel slot filling and other bulk tasks.
  • Semantic retrieval with reranker — Content is filtered first by embeddings (top-N), then re-ordered by a dedicated reranker (top-K). This noticeably improves match quality compared with embedding-only retrieval.
  • Material-grounded content — Slide texts are derived from typed inventory entries with source references. Gaps are surfaced as visible placeholders rather than filled in by the model.
  • Large layout catalogue with recommendation — Over thirty slide layouts are available (text, comparisons, processes, cycles, pyramids, matrices, hierarchies, metrics). A layout advisor selects the appropriate layout for each slide from the full catalogue.
  • Iterative detail editing in chat — Once the initial deck is generated, the application accepts free-text instructions, classifies them into one of 28 action types, and executes only the corresponding mini-operation. A classification check intervenes when the input is ambiguous and asks back instead.
  • Versioning with undo — Every modifying operation produces a version snapshot; the last 15 states can be reverted via undo.
  • Connection to four model services — institution-internal LLM infrastructure: a primary language model, a fast language model, an embedder, and a reranker.
  • Template-faithful export — A custom PPTX or POTX template is preserved; master layouts, fonts, and colours remain intact. Additional exports as DOCX and Markdown are available.