Refactor LLM interaction to use structured JSON filters

### 🧠 Refactor LLM to output structured filters (not raw Python)
```

Currently, LLM returns raw Python strings that are `exec()`-ed.

#### 💡 Advanced Fix:

* Prompt LLM to return structured JSON like:

```json
{ "column": "Date", "operator": "<", "value": "2024-01-01" }
```

* Your backend parses this and uses `pandas.query()` securely.

✅ This makes the system **robust, safe, and testable**.