Salim Laimeche

Jina AI was the only provider offering true multimodal embeddings — text and images in the same vector space — via a simple API. Our platform handles PDFs with figures, PPTX slides, and scanned documents, so cross-modal search (query text, retrieve images and vice versa) was non-negotiable. We evaluated OpenAI embeddings (text-only), Cohere (no multimodal), and local models like CLIP (operational overhead). Jina's CLIP v2 gave us 1024-dim multimodal vectors with a single API call and solid reranking on top. Clean DX, great performance.

Stripe was the no-brainer for SaaS billing in Europe. We needed subscription management with multiple tiers (Free to Enterprise), usage-based limits, and webhook-driven plan enforcement — Stripe handles all of it. The developer experience is unmatched: great docs, predictable webhook events, and a test mode that actually works. We considered LemonSqueezy (simpler but less flexible for B2B) and Paddle (better for tax but weaker API). For a platform selling to both devs and enterprises across EU, Stripe was the safest bet.

We needed auth that handles both individual users and multi-tenant organizations out of the box. Clerk gave us JWT-based auth, org management, RBAC, and webhook sync — all with a clean React SDK and middleware for Next.js. We evaluated Auth0 (too complex for our stage), Supabase Auth (no native org support), and NextAuth (too much DIY). Clerk let us ship secure multi-tenant auth in days instead of weeks, so we could focus on the RAG pipeline instead of reinventing session management.

Speed, simplicity, and fewer moving parts. Bun replaced Node + tsx + Jest + dotenv in one binary — our backend runs TypeScript natively with zero config. The built-in test runner, auto-loaded .env files, and native SQLite support removed a lot of tooling overhead. Cold starts are noticeably faster, which matters for our RAG pipeline where every millisecond counts. We switched from Node early on and never looked back.

Next.js with App Router gave us the best balance between developer experience and production readiness. Server components let us keep the dashboard fast while handling heavy data rendering (collection analytics, RAG config panels, workflow builder). The middleware layer was perfect for Clerk auth + i18n routing. We considered Remix and SvelteKit, but the React 19 ecosystem (assistant-ui for chat, Shadcn UI, Recharts) made Next.js the obvious choice for shipping a complex B2B dashboard quickly.

LangGraph (by LangChain) powers our agent orchestration layer. We evaluated building a custom state machine, but LangGraph's approach to cyclic graphs with built-in tool calling and checkpointing saved us months. The ability to define complex agent workflows with conditions, branches, and human-in-the-loop — while staying in TypeScript — was the deciding factor. It's the backbone behind our visual workflow builder.