-
-
Cani Agent Interface — Sleek glassmorphism UI showing system metrics, active session state, and policy limits.
-
Autonomous Approval — System verifying prompt parameters ($45.00) and authorizing PayPal API sandbox creation in real time.
-
Safeguard Interception — Cani identifies an over-budget request ($299.99) and instantly blocks execution to enforce policy limits.
Inspiration
With the rapid growth of autonomous AI agents interacting on behalf of humans, Agentic Commerce is becoming the next frontier in online transactions. However, giving an AI direct access to financial accounts introduces risks such as unintended purchases, hidden fees, and fraudulent checkouts.
We were inspired to build Cani, a dedicated AI safeguard layer that acts as an intelligent intermediary. It allows users to safely delegate shopping tasks to AI while ensuring every PayPal transaction undergoes real-time risk assessment, budget verification, and authorization checks.
What it does
Cani - Agentic Commerce & Autonomous Payment Safeguard Bot is an end-to-end AI purchasing assistant designed to make autonomous transactions safe, transparent, and user-controlled:
- Intent & Product Verification: Evaluates product details, seller reputation, and transaction parameters before initiating any payment action.
- Autonomous Budget Safeguard: Checks user-defined budget rules and spending caps using deterministic AI logic to prevent overspending.
- PayPal Integration: Dynamically generates and executes sandbox checkout orders via the PayPal Orders & Payments API.
- Auditable Log & Risk Scoring: Generates plain-language decision logs detailing why a transaction was approved or flagged.
How we built it
- Framework & AI Logic: Built using Python and LLM function calling to parse shopping requests, determine risk scores, and manage decision loops.
- Payment Architecture: Integrates directly with the PayPal Developer Sandbox REST API for handling order creation, authorization, and capture workflows.
- Frontend Dashboard: Designed a lightweight interactive UI using Streamlit to display live execution logs, budget meters, and transaction states.
- Database & Rule Engine: Uses SQLite to store real-time user policy configurations, token limits, and past transaction logs.
Challenges we ran into
- Handling Agent Determinism in Financial Decisions: Ensuring the AI strictly adheres to quantitative spending boundaries rather than hallucinating approvals required strict schema validation and fallback checks.
- Managing Sandbox Authorization Flows: Seamlessly orchestrating real-time API authentication and token management between autonomous agent execution loops and PayPal's sandbox endpoints.
Accomplishments that we're proud of
- Built a complete working prototype that successfully executes end-to-end PayPal sandbox transactions triggered purely by natural language intent.
- Created a transparent risk-scoring engine that provides clear explanations for every financial action taken.
What we learned
- Gained deep technical insights into constructing secure API integrations for AI agents handling real-time financial execution.
- Learned how to structure prompt boundaries and function-calling schemas to minimize financial automation risks.
What's next for Cani
- Multi-Currency & Subscription Safeguards: Expanding support for recurring payments and international transaction verification.
- Biometric Step-Up Authorization: Adding optional push-notification approval prompts for transactions exceeding dynamic threshold limits. *
