Ship your AI agent SaaS
this weekend.

Looked at Django (sync ORM, slower iteration), Flask (no async, no types, you build everything around it), Fastify (Node fights LLM Python libs every step). FastAPI is the only mainstream framework that delivers typed Python with async out of the box. Pydantic v2 is 5-50x faster than v1 and is what LangChain and the OpenAI SDK already speak natively.

SvelteKit and Remix have nicer DX but smaller pools of AI-app component libs. Vue / Nuxt loses the React landgrab in shadcn-style primitives. Next.js 15 + React 19 give server components (smaller bundles), Suspense streaming (token-by-token feels native), and the biggest selection of pre-built UI to drop in. App Router means marketing and product live behind one server.

Plain CSS-in-JS spirals into design debt at any scale. Bootstrap looks like Bootstrap. Tailwind 4 finally has CSS vars first-class, so a brand swap is one block in globals.css. We use shadcn primitives for Dialog, Select, Dropdown, Popover - no need to roll our own when the patterns are this well-trodden.

SQLite is fine for hobby projects, dies at multi-tenant. MongoDB makes row-level isolation messy and the agent has heavy relational reads. Postgres with explicit org_id columns is the boring choice that scales to billions of rows. SQLAlchemy 2.0's typed mapper is the closest Python comes to Rails-grade migrations and introspection. asyncpg is roughly 3x faster than psycopg2 for our read-heavy workload.

Looked at RQ (too simple, no schedules), Dramatiq (smaller community, fewer integrations), Taskiq (still young). Celery has been the production default for 15+ years - boring, stable, battle-tested. fastapi-cache2 gives us per-route caching with TTLs in three lines. Redis pulls double duty as cache + broker so the infra stays small.

Pydantic AI is cleaner but the pre-built tool ecosystem is shallow. CrewAI is great for multi-agent orchestration but overkill for a chat product. The Vercel AI SDK locks you to Node + serverless. LangChain has the deepest tool ecosystem and integrates with LangSmith out of the box. LangGraph handles state that LangChain alone can't. The agent module is isolated behind a single interface, swap runtimes if you want.

Pinecone is hosted-only and per-vector pricing punishes you at scale. Qdrant is great but smaller team. Chroma is fine for prototypes but slows past a few million vectors. Weaviate is Java-shaped and adds operational weight. Milvus has been proven at Zilliz, Bilibili and Walmart. Self-hostable via Docker Compose so your vector data never leaves your stack, and the hybrid retrieval beats pure-vector on factual queries.

Single-parser strategies fail on the long tail of file formats. liteparse handles ~80% of plain documents in milliseconds and zero API cost. LlamaParse via llama-cloud is the cleanest answer for multi-column PDFs, scanned tables, and messy academic papers - we route to it only when liteparse confidence drops. python-docx parses .docx natively (faster than going through PDF). pymupdf handles PDF text + layout. For images embedded in docs, we send them to a vision LLM, store the description in the index, and the agent answers "what's in the chart on page 4?" without OCR gymnastics.

We tried building billing from scratch once - lost two months to refund edge cases, EU VAT and SCA. Stripe is the only billing layer worth wiring for a serious SaaS. Customer Portal solves the cancel-flow regulatorily in every market we know. We add token-level credit metering on top of subscriptions because per-LLM-call cost variance is too high to bill per seat alone.

Auth0 / Clerk / Supabase Auth are great if you want a vendor lock and a per-MAU bill. Self-hosted auth on JWT + bcrypt is the boring path that costs nothing and ports anywhere. Authlib gives us OAuth provider integration without writing the dance ourselves. Session table in PostgreSQL means revocation is one SQL query, not a vendor support ticket.

Kubernetes is overkill for the first three years - Docker Compose runs on any VPS for $5/mo. Sentry is the established default for app errors; LangSmith from LangChain's own team is the only agent-specific observability that doesn't require building yourself. Nginx reverse-proxies, terminates SSL, and handles the WebSocket upgrade headers without configuration drama.