AI Capstone: your code drives an on-device LLM

This is the capstone, and you are the architect. The whole track has been building to one idea: your Python orchestrates the model. The LLM is just one component you call -> your code decides what data it sees, how the request is shaped, and what happens when no model is loaded. You will build that orchestration from a near-blank file.

Every real AI feature is the same loop. Your code does the reliable work, the model does the wording:

• Compute the facts in pandas. Code is exact at arithmetic; a model is not. So the row count, total, average, and top category come from a DataFrame, never from the LLM. This is the only way to avoid hallucinated numbers.
• Shape the request yourself. A reusable prompt template keeps the system instruction (who the model should be) separate from the user message (what it should do), then assembles them into one string that ends on a turn marker so the model knows where to start writing.
• Always have an offline path. The on-device model needs WebGPU and is non-deterministic, so you also write a rule-based generator that turns the same facts into a sentence. The feature works even with no model at all.

What you will build

Two functions, from scratch. build_prompt(system, user) returns one clean string that contains both parts and ends on Assistant:. insight(df) computes the stats on an embedded spending table and returns a plain-English line that names your biggest category and the total. A small spend_stats(df) helper does the pandas work so both the report and the insight read from the same numbers.

The AI step (try it in the app)

Floati runs its model client-side with WebLLM on WebGPU -> no account, no key, no cloud, your prompts never leave the machine. The app wires a single shared engine onto window._floatiTutor. Once you have built build_prompt, this is the real call (it lives outside the grader because it needs a GPU and is non-deterministic):

// Floati already created window._floatiTutor (a WebLLM engine).
const prompt = build_prompt("You explain budgets in one warm sentence.", report);
const reply = await window._floatiTutor.chat.completions.create({
  messages: [{ role: "user", content: prompt }],
  temperature: 0.5, max_tokens: 120,
});
console.log(reply.choices[0].message.content);

Notice the model only ever sees the report string your Python produced -> the numbers are already correct before the model speaks. Press Run to see your offline pipeline print its stats and its rule-based insight; that printed report is exactly what you would hand to the model above. The grader checks your prompt shape and your deterministic insight, never any model output.