Pickle is an AI-powered agent designed to function as a personal assistant that performs real-world tasks such as booking travel, scheduling social plans, making phone calls, and executing purchases. The service emphasizes a "privacy-first" architecture, positioning the AI as a secure intermediary between the user and external digital services.

Core Functionality and Workflow

The agent operates through an intent-driven interface where users delegate multi-step actions. The core methodology relies on a strict separation between user-sensitive data and the AI processing layer. By integrating directly with web and application interfaces, Pickle automates interactions that typically require human manual input.

Privacy and Security Architecture

Pickle utilizes a layered security model to ensure user data remains protected during AI operations:

1. Vault Encryption: Sensitive credentials, such as passwords and credit card numbers, are encrypted using a cryptographic key generated locally on the user's mobile device ($K_{local}$). This key is never transmitted to or stored on the company's servers.
2. Isolated Execution: For high-stakes operations (e.g., payments), the system employs a "sealed vault" mechanism. When an action requires authentication or payment, the system only decrypts the specific field value ($V_{field}$) needed for that transaction:

$$V_{field} = Decrypt(C_{vault}, K_{local})$$
where $C_{vault}$ is the encrypted ciphertext. This ensures the AI processor receives only the necessary token for the current task without gaining persistent access to the full set of user credentials.

3. Data Isolation: Conversations and user memories are processed independently from credentials. These are encrypted using a separate server-side key ($K_{server}$), ensuring that even if one layer is compromised, credentials remain inaccessible to the AI model or unauthorized staff.
4. Human-in-the-Loop Verification: To mitigate risks associated with autonomous actions, the system implements a verification protocol for irreversible transactions. If a transaction cost $C_{total}$ exceeds a threshold $T$ or changes in real-time ($C_{new} > C_{original}$), the agent is programmatically forced to halt and seek explicit user authorization before execution.

Operational Transparency

Pickle maintains a strict policy against using personal user data, conversations, or content for the training of Large Language Models (LLMs). While external AI providers may temporarily cache data for abuse prevention, the internal policy dictates that the platform does not sell information to third parties for advertising. The platform further logs all instances of authorized staff access, ensuring an audit trail for internal maintenance or legal inquiries, while strictly masking sensitive vault data even from internal personnel.