cd ../case-studies
~/case-studies/slidecrafter-ai.md
Architectural Deep-Dive

SlideCrafter AI

A local-first presentation compiler that turns probabilistic model output into structured, editable decks.

SlideCrafter AI screenshot

structured AI compilation

system flow

Probabilistic generation ends before deterministic validation and file compilation begin.

Sources
notes + documents
→
Model
structured IR
→
Validator
repair constraints
→
PptxGenJS
editable deck

Product context

SlideCrafter AI explores how to turn research material into editable presentation decks without asking a language model to directly control every pixel. The product treats presentation generation as a compiler pipeline.

Pipeline

  1. The model turns source context into a structured intermediate representation.
  2. The application validates the structure, theme primitives, and layout constraints.
  3. PptxGenJS compiles the validated representation into an editable .pptx file.
  4. Invalid output can enter a bounded repair loop instead of silently producing a broken deck.

Engineering decisions

  • React, TypeScript, and Vite keep the application fast to iterate on.
  • Structured instructions make model output inspectable and replaceable.
  • Local-first processing keeps compilation close to the user and reduces unnecessary data movement.
  • Deterministic compilation separates probabilistic generation from file creation.

Evidence and limits

The public deployment demonstrates the compiler direction and editable deck output. I do not claim universal layout correctness or a measured monthly user count. The useful proof is the architecture: model output is constrained, validated, and compiled rather than trusted blindly.