Mr Haseeb outlines AI stack

Haseeb Qureshi — the Paradigm partner who co-founded Basis and knows VC + crypto inside out — just dropped a dead-simple AI app stack on X. Founders waste weeks googling "best AI backend 2025." He fixed that. This playbook gets you from zero to prototype covering web, iOS, Android in days, not months. It handles 90% of what startups actually ship today. ### Who is Haseeb and why listen? Haseeb's not some rando dev influencer. He's a poker pro turned crypto VC who built Basis (AI trading infra) and invests in stacks like this daily. His take: most AI "tutorials" are toy demos. Real startups need production-grade, cross-platform from day one. This stack delivers — battle-tested for scale. ### What's the frontend stack? Next.js for web. It's React with file-based routing, server-side rendering baked in — zero config headaches. Pairs perfectly with AI because you can stream LLM responses server-side, no client hacks needed. React Native for mobile. One codebase deploys to iOS + Android. No Swift/Kotlin boilerplate. Both share React DNA, so components reuse across platforms. Dead simple for AI apps needing real-time chat or agent UIs. ### Backend and auth — why these picks? Node.js runtime. Fast for I/O-heavy AI calls (LLMs are all network). Supabase for everything else — Postgres DB, auth, realtime subs, edge functions. It's Firebase but open-source, SQL-native. No vendor lock. Auth is OAuth/JWT out of box. Scales to millions without AWS sprawl. Haseeb calls it "the Vercel of backends." Turns out, 80% of AI startups undervalue auth — this nails it first. ### The AI layer — where magic happens? OpenAI or Claude APIs as the brain. Pick based on use case — GPT-4o for vision/speed, Claude for reasoning depth. LangChain.js orchestrates: chains prompts, handles tools, retries flakes. Not mandatory but cuts boilerplate 10x. For RAG (retrieval-augmented generation), Pinecone vector DB if you want managed simplicity, or pgvector in Supabase Postgres for free/embedded. Query your docs, not hallucinate. ### Why 90% coverage? This hits web/mobile/backend/AI/RAG/auth/DB/realtime. Missing? Heavy ML training (use HF/Replicate), ultra-custom infra (rare for v1). But for chatbots, agents, analytics apps — it's 90%. No Kubernetes yaml wars. Deploy to Vercel (web/mobile) + Supabase (backend) in hours. Cost: ~$50/mo at startup scale. ### How does it all wire together? Example flow: User chats in Next.js app → Node endpoint hits LangChain → Chain pulls RAG from pgvector → Claude generates → Stream back via Server-Sent Events. React Native mirrors the UI. Supabase syncs user state realtime. Secure, scalable, cheap. Haseeb's post links boilerplate repos — fork and ship. ### What's the catch? Vector costs add up at 1M+ queries/mo — Pinecone bills per vector. LangChain abstracts well but debug abstractions when things flake. Mobile AI inference? Still client-side limited; this offloads to cloud. Test on device early. ### Why now? AI startups exploded post-2024 — but tooling lagged. Founders still duct-tape Supabase + Vercel + OpenAI manually. Haseeb's stack is the 2025 default. 71 views so far, but it'll compound as indie hackers clone it. If you're bootstrapping AI, this is your weekend project. Bottom line: Copy this stack. Prototype faster than competitors fumbling stacks. Haseeb just saved the next 1,000 AI founders six figures in engineer time. ``` (Word count: 528)