How to ship an AI MVP in 2 weeks

Write a one-page spec: the user, the single core job the AI does, the input, the output, the metric for "good". Pick the model and framework (Claude Sonnet 4.6 or GPT-5, Vercel AI SDK, Next.js 15). Clone a starter, get a hello-world completion working locally, push to Vercel. Build the first 30 examples of your eval set by hand.

Build the end-to-end happy path in under 500 lines. Single prompt, single call, no retries, no streaming, no auth. Ship it as a one-page app with a textbox and an output. Invite 3 test users. Capture every input and output in a database table. If the core thing is not useful in 3 days, scope is wrong and you fix that before adding anything.

Add retrieval if you need domain knowledge (pgvector + hybrid search + reranking). Wire LangSmith or Braintrust and run your eval set on every prompt change. Add tools if you need actions (Vercel AI SDK tool calling, 4 to 8 tools max). Iterate prompts against evals, not vibes. By day 9 the output should hit your success metric on 70 percent plus of eval cases.

Add auth (Clerk or Supabase), rate limiting (Upstash), prompt caching (cache_control on Anthropic, automatic on OpenAI above 1,024 tokens), streaming UX, error states, and a feedback thumbs button that writes to your eval set. Add logging and a trace per session. Run a load test with 50 concurrent requests.

Deploy to production. Invite 20 to 50 real users. Watch the traces live for the first 24 hours. Pick the 5 worst outputs, add them to the eval set, fix the prompt, redeploy. By end of day 14 you have a system with real users, real data, real metrics, and a clear list of what to build next.

Do not fine-tune in week 1

Fine-tuning requires labeled data you do not yet have and locks you to a model version. Prompt engineering, RAG, and tool use get you 90 percent of the benefit with 10 percent of the effort.

Do not build a custom vector DB

You are not Pinecone. pgvector or Turbopuffer works for the first year. Custom shards, custom indexing, custom reranking pipelines are week-6 problems, not week-1 problems.

Do not chase every eval metric

Pick 3 to 5 metrics tied to user outcome (task success rate, citation accuracy, refusal rate, cost per call). Ignore the rest until you are optimizing for scale.

Do not wire up 12 models

Model routing is a week-8 optimization. In week 1, pick one strong model for the main path and one small model for cheap subtasks. Ship. Measure. Then route.

Do not over-architect the agent loop

One system prompt, 4 to 8 tools, a 10-step max, a cost cap. That is enough for 80 percent of production agents. Multi-agent orchestration, memory systems, and reflection loops are upgrades, not starting points.

Do not skip prompt caching

Anthropic prompt caching cuts costs 70 to 90 percent on repeat prefixes. It takes 30 minutes to set up. Skipping it in week 2 means paying a 5x markup on every call for the next 6 months.

Is 2 weeks realistic for a production AI MVP?

Yes, if the scope is narrow and one person owns the decisions. We have shipped RAG apps, agentic support bots, and AI-powered internal tools in 10 to 14 working days with teams of 1 to 3. The key constraint is scope, not staffing.

What stack do you recommend?

Claude Sonnet 4.6 or GPT-5 for reasoning, Haiku 4.5 for cheap tasks, Vercel AI SDK for glue code, Next.js 15 on Vercel for the app, Postgres plus pgvector for storage and retrieval, LangSmith or Braintrust for evals, and Supabase or Clerk for auth. Resist the urge to add anything else.

How much does a 2-week AI MVP cost?

Self-built by a capable founder-engineer: 10 to 14 working days plus $200 to $1,000 in API and infra spend. Hired out: $25k to $50k for a specialist team. The model cost alone during development is usually under $500 if you cache prompts correctly.

Should I fine-tune a model for an MVP?

No. Fine-tuning in week 1 is almost always a mistake. Prompt engineering, RAG, and tool use get you further faster. Fine-tuning becomes worth it only after you have real user data, stable requirements, and a reason the prompt approach is topping out.

Do I need LangChain or LangGraph for an MVP?

Usually no for the first 2 weeks. Vercel AI SDK with native tool calling is lighter and ships faster. Reach for LangGraph when you need multi-agent orchestration, long-running workflows, or stateful graphs. Start simple and upgrade when a real pain appears.

What about user auth, billing, and infrastructure?

Clerk or Supabase Auth handles users in a few hours. Stripe for billing. Vercel for hosting. Upstash Redis for rate limiting. Do not build these. Buy them. Every hour you save on plumbing is an hour spent on the AI surface that actually differentiates you.

Do I really need evals this early?

Yes. A 50-example eval set from week 1 pays for itself by day 7 when you are tweaking prompts and need to know if you are improving. Use LangSmith or Braintrust. It takes 2 hours to set up and saves weeks of regression hunting later.

When is 2 weeks not enough?

When the app needs enterprise SSO, HIPAA or SOC 2, complex approvals, multi-tenant data isolation, or deep integration with legacy systems. Also when the core AI problem is genuinely novel research. Be honest about which category you are in and scope accordingly.