AlgoJinn

Crypto trading interfaces are dense and intimidating. The client wanted people to act in plain language — “sell half my ETH if it drops below 3,000” — while keeping real money completely safe.

A Gemini assistant interprets intent and proposes a structured trade; the backend owns everything that touches money. The model never executes — it suggests a typed call, and the server validates, runs risk and balance checks, and only then places the order.

A Python/MySQL backend exposes trade and market functions that Gemini maps natural language onto via function calling. Flutter clients for mobile and Windows ship from one codebase, structured with the BLoC and Repository patterns, and payments run through RazorPay, Stripe, PhonePe, and CCAvenue.

Making an LLM safe around money was the whole game: validate and clamp every argument, confirm destructive actions, and fail closed whenever intent is ambiguous rather than guessing.

Keeping mobile and Windows in lockstep from a single Flutter codebase took disciplined state management.

Treat model output as a request, never a command. The guardrails — confirm, clamp, fail-closed — are the actual product; the language interface is just the front door.