KumoKodo.ai Case Study
Building LisAI
How we built an AI-powered social media content creation platform using Claude Sonnet 4, Google Imagen 3, and Google Veo 2 to generate, schedule, and publish across 6 platforms from a single command.
The Challenge
Businesses and creators struggle to maintain a consistent social media presence across multiple platforms. Creating platform-specific content, generating visuals, scheduling posts, and maintaining brand consistency is a full-time job — or an expensive team.
- Creating unique content for 6+ platforms is time-consuming and repetitive
- Hiring a social media team costs $5,000–$15,000/month for most businesses
- Maintaining brand voice consistency across platforms is nearly impossible manually
- Generating professional images and videos requires separate tools and subscriptions
- Scheduling and publishing across platforms means juggling multiple dashboards
- No existing tool combines AI text, image, and video generation with direct publishing
Our Solution
LisAI combines three best-in-class AI models into a single platform that generates text, images, and video content — then schedules and publishes it directly to your connected social media accounts.
AI Content Generation (Claude Sonnet 4)
Brand-aware, platform-specific content generation powered by Anthropic Claude Sonnet 4. Each post is tailored to the platform's format, character limits, and best practices. Supports SEO keyword injection, A/B/C variant generation, and content regeneration with custom user feedback.
AI Image Generation (Google Imagen 3)
Automatically generates professional-quality images from text prompts using Google Imagen 3.0. Images are uploaded to Google Cloud Storage with signed URLs, providing reliable hosting for social media publishing.
AI Video Generation (Google Veo 2)
Creates short-form video content using Google Veo 2 with configurable aspect ratios and durations. Uses an async long-running operation pattern with plan-aware duration caps (5–8 seconds) for optimized usage.
Smart Scheduling (GCP Cloud Tasks)
Schedule individual posts or bulk-schedule entire content plans. GCP Cloud Tasks handles the publish queue with shared-secret authentication and a daily Vercel cron fallback for posts beyond the 30-day window.
Technical Deep Dive
Backend Architecture
Built on Next.js 15 with the App Router, using route groups to separate marketing pages from authenticated app routes. Drizzle ORM provides type-safe queries against Neon serverless PostgreSQL.
- 30+ API endpoints across 8 resource groups
- Zod validation on all request payloads
- server-only module boundary to prevent client leaks
- JWT sessions via Auth.js v5 with Drizzle adapter
AI Integration Layer
A best-of-breed AI mix — Claude for text, Imagen for images, Veo for video — each chosen for its strengths. Brand context is injected into every generation prompt.
- 1.Claude Sonnet 4: Platform-specific content with brand context injection
- 2.Imagen 3.0: Text-to-image with exponential backoff retry logic
- 3.Veo 2: Async long-running video generation with polling pattern
- 4.GCS upload pipeline for persistent image and video storage
Frontend Engineering
React 19 with Tailwind CSS 4 powers the UI. The app includes 13 distinct pages from content generation to calendar scheduling, all with responsive design.
- Monthly calendar grid with platform color-coding
- Natural language content plan creation
- Brand management with website scanner and document extractor
- Usage meters and billing portal integration
Database Design
Neon serverless PostgreSQL with Drizzle ORM. 12 tables across auth, application, and billing domains with 7 custom enums for type safety.
- Auth tables: users, accounts, sessions, verification tokens
- Application: social connections, brands, content plans, generated content, media
- Billing: subscriptions, usage records, addon purchases
- Content plans stored as JSON arrays for flexible planning
Key Features
Brand Intelligence
AI-powered brand extraction from websites and documents. Claude analyzes your brand's industry, audience, tone, and messaging, then injects it into every piece of content.
- Website scanning with Claude extraction
- Document text analysis
- Brand context in all prompts
- Multi-brand management
Batch Content Planning
Create content plans with natural language prompts. Choose from 3 approval modes: Plan First, Autonomous, or Manual Review.
- Natural language plan creation
- SEO keyword injection
- A/B/C variant generation
- Approve & Generate All
Multi-Platform OAuth
Connect LinkedIn, Facebook, Instagram, and TikTok with unified OAuth flows. Automatic token refresh with 5-minute expiry buffer.
- LinkedIn, Facebook, Instagram, TikTok
- Unified dynamic route pattern
- Auto token refresh
- Facebook Page selection
Calendar View
Monthly calendar grid showing all scheduled and posted content, color-coded by platform with filtering and inline actions.
- Platform color-coding
- Status filtering
- Approve & publish from calendar
- Post detail sidebar
Billing & Subscriptions
4-tier SaaS model from Free to Business ($99/mo) with Stripe integration, usage metering, and add-on packs for extra capacity.
- Per-action usage metering
- Add-on packs (+Videos, +Posts)
- Feature gating by tier
- Stripe billing portal
Post Management
Full lifecycle tracking from draft to published. Regenerate content with custom feedback, track scheduling timestamps, and monitor publish errors.
- 6-stage status tracking
- Content regeneration
- SEO keyword storage
- Video content support
Results & Metrics
Lessons Learned
Best-of-breed AI beats one-size-fits-all
Using Claude for text, Imagen for images, and Veo for video — each chosen for its strengths — delivered far better results than any single model could. The integration complexity pays for itself in output quality.
Design for async from the start
Video generation takes 30–90 seconds. Building the async polling pattern early (POST to start, GET to check) meant we could extend it to batch content generation without refactoring.
Brand context transforms AI output
Injecting extracted brand data (industry, audience, tone, key messages) into every Claude prompt was the difference between generic AI content and content that sounds like the brand.
Usage-based billing needs upfront architecture
Check limits before generation, record usage after success, handle add-on packs extending quotas. Getting this right early prevented billing disputes and made tier upgrades seamless.
About KumoKodo.ai
We specialize in building AI-powered SaaS applications using modern cloud-native technologies. Our team combines expertise in machine learning, full-stack development, and user experience design to create products that solve real business problems.
Want to Try LisAI?
Start creating AI-powered social media content today. Or contact us to discuss building a similar platform for your business.