Gemma SlideCast: A 4-Agent Autonomous Presentation & Document Engine

This is a submission for the Gemma 4 Challenge: Build with Gemma 4

What I Built

GemmaSlide is an autonomous, multi-agent slide presentation & document generation engine. It turns long, unstructured prose into beautiful, print-ready visual slide decks (matching aesthetics like NotebookLM or Minimal Modern styles) in seconds.

Traditional document summaries are text-heavy and dull. GemmaSlide solves this by orchestrating four specialized, cooperative AI agents in a pipeline that handles structure, design, quality audit, and unified output formatting:

1. The Planner Node: Dissects raw prose and maps out a cohesive presentation storyboard.
2. The Designer Node: Translates visual style sheets directly into beautiful modular HTML layouts.
3. The Critic Node: Audits contrast ratios, margins, and readability, returning corrective styling scores.
4. The Assembler Node: Stitches pages together, inserting smart cover pages, tables of contents, and print pages under unified CSS templates.

---

Repo

• Git Hub Repo Link: https://github.com/Jagadeesh961982/dynamicpdfgenerator

---

Code

• Backend & Frontend Repository: [Insert Link to your GitHub Repository]
• Core Stacks:
  • Backend: Python + FastAPI (with server-sent events for streaming agent states)
  • Frontend: React + TypeScript + Tailwind CSS (with Recharts for telemetry)
  • Orchestration: Multi-Agent Pipeline utilizing the official Google GenAI SDK.

---

How I Used Gemma 4

Our application relies on a configurable pipeline topology where users choose the exact model for each of the four roles. This highlights model trade-offs between speed, reasoning accuracy, and styling compliance.

Here is how each model was choose to power our workspace:

🚀 Which Model is Right for Us? Our Topology Selection

• Gemma 4 (9B - SOTA Instruct): Chosen for the Planner and Designer roles. Its state-of-the-art instruction tuning delivers 2.5x higher structural alignment than similar-tier models. It consistently produces raw JSON slide storyboards and clean HTML layouts without breaking or outputting markdown junk.
• Gemma 2 (27B - Heavy Reasoning): Selected as our Critic. Reviewing contrast ratios, reading text configurations, and outputting highly specific design audits requires deep intelligence. Gemma 2's 27B parameter weight was the clear winner for safety assessment.
• Gemma 2 (9B): Chosen as our super agile Assembler to quickly handle structural page-break compilation and unify final document styling at incredibly low latency.

📊 Latency Telemetry

Our frontend tracks precise diagnostic telemetry for every run on a dynamic Recharts bar chart. Running Gemma 4 (9B) for creative generation alongside Gemma 2 (27B) for critique creates an optimized blend of execution speed and human-quality structural design!