Inspiration
Every team already has the raw ingredients for great content: brand guidelines, product screenshots, founder notes, launch decks, past campaigns. But these assets are scattered across dozens of Google Drive folders, making content creation slow, inconsistent, and impossible to scale. We watched teams spend hours manually assembling context before they could even start writing a caption. We wanted to build something that turns that fragmented knowledge into a shared creative intelligence layer, so any team member can produce on-brand content for any platform in minutes.
What it does
WiseStory is a multi-tenant, AI-powered creative storytelling platform. You connect your company's Google Drive, and WiseStory ingests your brand materials (guidelines, images, tone docs, campaign references) to build a workspace-level knowledge base. Then, through a multi-turn briefing chat, a creative director agent collaborates with you to generate platform-ready content: Instagram posts and carousels, Reels, TikTok scripts, YouTube Shorts, long-form video scripts, and full multi-platform campaign packages.
The key difference: outputs are interleaved multimodal. Text, generated images, storyboards, scripts, come together in one cohesive creative package, not bolted together after the fact. Every piece of content references the exact source materials that shaped it, so you always know where the ideas came from.
How we built it
We built WiseStory as a TypeScript monorepo using Turborepo with pnpm workspaces. The frontend is a Next.js 16 app with React 19, Tailwind CSS 4, and shadcn/ui components, authenticated with Better Auth via Google OAuth.
The agent service runs on Hono and is orchestrated with the Google Agent Development Kit (ADK v0.5). It uses Gemini 2.5 Flash Image for interleaved text and image generation in a single stream, and Gemini Embedding 2 Preview for multimodal embeddings that power workspace-level RAG across documents, images, and PDFs. Embeddings are stored as 768-dimensional vectors in PostgreSQL 16 with the pgvector extension, enabling semantic similarity search scoped to each workspace.
The entire platform runs on Google Cloud: Cloud Run for containerized services, Cloud SQL for the database, Secret Manager for credentials, and Artifact Registry for container images.
Challenges we ran into
Getting interleaved multimodal output right was the hardest part. Most AI workflows generate text first, then images separately, and leave the user to assemble everything. We needed Gemini to produce text and images together in one coherent stream, which required careful prompt engineering and output structuring through ADK.
Workspace-level RAG with tenant isolation was another challenge. Every retrieval query must be scoped to a single workspace, and project-level context needs to layer on top of workspace knowledge without leaking across boundaries. Designing the embedding pipeline to handle documents, images, and PDFs through a single multimodal embedding model (Gemini Embedding 2 Preview) took iteration to get right.
Accomplishments that we're proud of
We built a fully deployed, multi-tenant SaaS product with real Google Drive integration, workspace-scoped RAG, and true interleaved multimodal generation. The platform generates complete creative packages (scripts, storyboards, images) in a single agent response, not as disconnected pieces stitched together.
We are also proud of the architecture. The monorepo shares one Prisma schema, one set of Zod contracts, and one auth model across all services. This kept us moving fast without schema drift or integration headaches.
What we learned
Multimodal embeddings are a game changer. Using Gemini Embedding 2 Preview to embed text, images, and PDFs into the same vector space makes retrieval dramatically simpler and more powerful than maintaining separate pipelines for each content type.
We also learned that media-type awareness matters more than we expected. An Instagram carousel has completely different structure, pacing, and visual needs compared to a YouTube Shorts script. Making media type a first-class concept in both the data model and the prompt architecture was essential for producing outputs that actually feel native to each platform.
What's next for WiseStory
We plan to add real-time Google Drive sync so workspace knowledge stays current without manual re-indexing. We want to expand to more content formats like LinkedIn posts, email sequences, and podcast outlines. We are also working on campaign version history so teams can track how their creative direction evolves over time, and a PDF/Markdown export flow for easy handoff to production teams. Longer term, we want to add collaborative workspaces with role-based access and direct social publishing integrations.
Log in or sign up for Devpost to join the conversation.