Building a multi-tenant AI-powered LinkedIn content platform for a marketing team that needed brand consistency across multiple workspaces.

PostPilot is an AI-powered LinkedIn post generator I built for Prompt (meetprompt.com), a marketing and communications company led by Paul. Their team was creating LinkedIn content manually — different people writing in different voices, no shared brand guidelines, no way to track what was working, and no central system for managing who writes what across multiple teams. I designed and built the entire platform end-to-end: a multi-workspace system where different teams operate independently while a super admin oversees everything from one dashboard. The system turns a brief into three distinct, on-brand LinkedIn post variations in under a minute using Claude AI, learns from approved posts to improve over time, and generates editorial-quality images using Google Imagen 4. Built with Next.js 16 (App Router), TypeScript, Supabase (Postgres + Auth + Storage + Row Level Security), Anthropic Claude API (Haiku + Sonnet), Google Imagen 4 for image generation, and Serper.dev for competitive intelligence.

Engineering a multi-tenant content platform with AI generation, brand voice learning, and real-time streaming — all with workspace-level isolation.

The core challenge was architectural: building a true multi-tenant system where every piece of data — generations, voice profiles, content strategies, approved posts, competitive intelligence — is isolated per workspace through Supabase Row Level Security, while giving a super admin full visibility across all workspaces from a single dashboard. On top of that, the AI generation pipeline needed to feel instant despite making multiple Claude API calls — so I built SSE streaming that delivers tokens to the UI as they arrive, with each of the three variations streaming independently. The image generation pipeline was another challenge: naive text-to-image prompts produced generic stock photo results, so I built a two-stage system where Claude Haiku first analyzes the post content like an art director and proposes a concrete visual scene, then Imagen 4 generates the actual image. The platform also needed invite-only access control with two-tier roles (platform-level super admin and workspace-level owner/admin/member), per-user voice document selection, a feedback loop where approved posts improve future generations, and rate limiting that works across both burst and daily windows.

A content platform that transformed how the team creates LinkedIn posts — structured, on-brand, and AI-powered.

The Prompt team went from scattered manual content creation to a structured, AI-powered workflow where every post stays on brand and every team member works from the same system. The multi-workspace architecture gave leadership full visibility while letting individual teams operate independently. The brand voice learning meant posts sounded like their team — not like AI. The super admin panel with golden badges, assign-to-workspace modals, and workspace-scoped user management gave Paul's team the control they needed to manage everything from one place.

Domain-driven multi-tenant architecture with streaming AI generation, intelligent image prompting, and a self-improving content loop.

Designed the system around a domain-driven folder structure with clean separation between UI (features/), business logic (domain/), and external service adapters (infrastructure/). Built 45 API endpoints covering generation, brand voice, content strategy, competitive intelligence, content calendar, admin operations, and audit trails. The generation pipeline builds prompts from multiple context sources — brand voice profile, content strategy pillars, approved post examples (few-shot learning with 90-day recency and quality weighting), and anti-AI-tell rules — then streams three distinct variations via SSE using a 5-format by 5-angle matrix that enforces genuine diversity between options.

Built the multi-tenant system with workspace-level RLS policies, a two-tier role system (platform roles for super admins, workspace roles for owners/admins/members), invite-only onboarding with configurable member caps, and per-user voice document selection so teammates choose their active voice independently. The image generation pipeline uses Claude Haiku as an art director to extract visual concepts, then Google Imagen 4 to generate editorial-quality photographs with a rolling history of 4 attempts per variation and automatic storage cleanup. Added session resilience (JSON 401 on API routes with global SWR redirect), audit trail logging, competitive post discovery via Serper, and a 13-tag feedback system across 6 dimensions that steers future generation quality. The full platform ships with 123 unit tests on the domain layer.