Teaching an LLM to trade: function calling with Gemini

"Sell half my ETH if it drops below 3,000." A sentence like that is trivial for a person and historically painful for software. Function calling is what closes the gap: the model doesn't execute anything — it decides which of your functions to call, and with what arguments.

Describe the tool, not the prompt

The model only needs a schema. Everything else — auth, risk checks, the actual order — stays in your code.

const placeOrder = {
  name: "place_order",
  description: "Place a crypto trade on the user's connected exchange.",
  parameters: {
    type: "object",
    properties: {
      symbol: { type: "string", description: "e.g. ETH, BTC" },
      side: { type: "string", enum: ["buy", "sell"] },
      quantity: { type: "number" },
      limitPrice: { type: "number", description: "Optional limit price" },
    },
    required: ["symbol", "side", "quantity"],
  },
};

The model proposes; your code disposes

Gemini returns a structured functionCall. I treat that as a request, not a command: validate the arguments, run risk and balance checks, and only then place the order. The LLM never touches the exchange directly.

Guardrails are the product

• Confirm destructive actions before executing them.
• Clamp quantities to the user's actual balance.
• Log every call with its arguments for an audit trail.
• Fail closed: if anything is ambiguous, ask instead of guessing.

The magic isn't that an LLM can trade. It's that natural language becomes a typed, validated call into code you already trust.