A shared, privacy-safe “credit score for returns.” Bastion plugs into any returns flow, runs a one-time KYC on high-risk events, issues a universal Bastion ID (UUID), and returns an actionable risk score + reasons - without sharing raw PII across merchants.
Return fraud silently drains e-commerce, costing retailers over $103 billion in 2024 alone (read more about it here)! False “package never arrived” claims, wardrobing, and empty-box returns cost far beyond merchandise; shipping, labor, and trust take the hit. We built Bastion to give developers and ops teams a drop-in, API-first shield that strengthens with every integration.
- Add enterprise-grade fraud checks to existing workflows in minutes
- Pull a user’s risk score and dispute-history signals (cross-platform, privacy-preserving)
- Automate decisions (approve/hold/manual review) via outbound webhooks
- Visualize most-disputed products & categories and risk distributions
- Drill into claim data (first-time vs. repeat, item category, etc.)
- Trigger - A return/dispute is initiated (e.g., “package not received”) on a merchant’s site
- Create Claim - Merchant calls
POST /claimswith event + minimal metadata - Selective KYC - If high-risk, Bastion triggers one-time KYC and issues a Bastion ID
- Risk Scoring - Bastion computes a risk score using:
- Behavioral & velocity signals (repeat patterns, dispute frequency, category spikes)
- Semantic similarity to historical fraud narratives (Cohere Rerank over a vector DB)
- Decisioning - Bastion returns dispute data and risk score based on company and customers
- Learning Loop - Merchant posts final outcome using inbound webhook; models & category baselines update continuously
- Backend: FastAPI (Python) services for claims, users, and analytics
- Database: Supabase (PostgreSQL) with strict schemas and row-level security
- Frontend: React (Vite) + TypeScript, modular UI packages shared across apps
- Pattern: Service-oriented modules with clear interfaces and typed DTOs
- Orchestrates claim intake, selective KYC, scoring, and decisioning
- Fraud detection engine: uses Cohere Rerank to compare new claims against a vector DB of historical fraud cases
- Produces a semantic relevance score that feeds the dynamic risk model
- Helps uncover non-obvious patterns and emerging fraud narratives
- Exposes clean REST endpoints (
/claims,/users,/analytics) with validation and audit logging
- Merchant Dashboard: BI-style views with interactive charts (Recharts)
- Explore fraud trends, risk distributions, and top disputed items/categories
- Drill into data (first-time vs. repeat claimants, by category, by time)
- Customer-Facing KYC Widget:
- Secure flow for completing verification
- Minimal friction; responsive UI that fits any modern checkout/returns page
- Outbound webhooks: push real-time decisions to merchant systems (e.g., “hold refund,” “manual review”)
- Inbound webhooks: merchants notify Bastion of final outcomes (accepted/denied)
- Creates a closed learning loop so feature weights and category baselines improve over time; no polling required
- Actionable Insights vs. User Privacy - Providing merchants with valuable fraud data without exposing a customer's PII (Personably Identifiable Information) or their shopping history with a competitor was a critical design challenge. We solved this by focusing on anonymized risk signals and aggregated behavioral patterns rather than raw personal data.
- Balancing AI Sensitivity - Fine-tuning our AI-powered risk model to minimize false positives (flagging a legitimate customer) while effectively catching sophisticated fraud was an iterative process. We focused on creating a scoring system that was firm but fair. Data Consistency Across Platforms: Architecting a system to handle data from multiple sources required robust validation and a clear source of truth. We designed the Bastion ID to be the central key that links disparate claims without creating data conflicts.
- Data Consistency Across Platforms - Architecting a system to handle data from multiple sources required robust validation and a clear source of truth. We designed the Bastion ID to be the central key that links disparate claims without creating data conflicts.
- Production-Grade KYC Integration - Our priority is to integrate a production grade KYC service like Clear or AiPrise's third-party KYC service to replace our own verification system. This will provide enterprise-grade identity verification, enhancing our fraud detection accuracy and ensuring compliance.
- Customizable Risk Engines - We plan to allow retailers to customize fraud detection parameters based on their specific industry (e.g., fashion vs. electronics) and risk tolerance, including configurable thresholds and custom scoring rules via the API.
- Risk Calculation & Enterprise Features - Our vision is to expand our capabilities with more sophisticated ML models to determine a more accurate risk factor. We also plan to build out enterprise-grade features, including customizable dashboards and personalized reporting to support the largest e-commerce platforms.