This repository contains the MACH X Hackathon Fraud Detection prototype—a fully orchestrated, AI-powered DMCA analysis workflow built using Conscia’s Hybrid AI Orchestration Engine. The project demonstrates how AI, data, and APIs can be orchestrated to automate fraud detection and takedown workflows with auditability, transparency, and human oversight.
- Purpose: Provides structured context about current manual processes
- Format: PDF documents, process guides, standard operating procedures
- Integration: Transformed into machine-readable context for LLM consumption
- Update Frequency: As processes change
- Purpose: Training data from 2–3 years of manual DMCA takedown history
- Content: URLs, violation types, success/failure patterns, resolution times
- Format: JSON export from Trello/Monday.com
- Volume: ~1,000+ historical cases
Used to train the model’s understanding of patterns and support retrieval workflows.
- OpenAI GPT-4
- Claude Sonnet
- Model-agnostic architecture (any LLM can be plugged in)
- Historical takedown data informs predictions
- DMCA documentation injected for structured reasoning
- Real-time classification using blended context
Purpose: Structured storage of detected violations and metadata
Schema: URLs, violation types, confidence scores, reasoning, status
Query Interface: Real-time APIs for retrieval and updates
DX Graph stores:
- Violating URL history
- Non-violating URL history
- Metadata + embeddings
- Full audit trace of all decisions
- Every decision has full lineage
- All inputs → transformations → outputs are recorded
- Complete historical logs for compliance
- LLM call history
- Orchestration events
- Error logs & retries
- Real-time monitoring
-
Input Sources
- Trello tickets supply new cases and reference data
- Slack accepts newly submitted suspicious URLs
-
Long-Term Memory (DX Graph)
- Stores violating & non-violating URL history
- Used to fetch contextual examples for the LLM
-
Page Scraping
- Submitted URLs are scraped for content
-
LLM Context Preparation
- Combines scraped content + memory examples + DMCA rules
-
LLM Execution
- Selected LLM classifies as violating / not violating with reasoning
-
Post-Processing
- Results returned to Slack for human review
- Confirmed violations added to DX Graph
-
Feedback Loop
- Slack reviewers approve or override
- Confirmed signals improve future model accuracy
- API & LLM orchestration
- Context injection & state management
- Data transformation
- Event-based triggers
- Webhooks for Slack notifications
- Human-in-the-loop approvals
No glue code required.
- Analysts receive alerts
- Can review, approve, override
- Feedback updates DX Graph memory
- Triggers: New violations, status changes, threshold alerts
- Destinations: Slack, external systems
- Payload: Structured violation data
- Reliability: Retries + failure handling
- Usage-based licensing
- Pricing based on monthly API call volume
- Prototype operates under 100K API calls/month
- Base license included
- Supports up to 10,000 records
- Enterprise tiers available
- Historical ingestion included
- Real-time inference billed at model provider passthrough cost
- Slack licenses assumed existing
- Implementation options:
- Self-directed
- MACH-aligned SI
- Conscia Professional Services
Special thanks to contributing companies, technologies, and BÆRSkin Tactical for supporting the MACH X Hackathon proof of concept.
Built by the Conscia-led MACH X Hackathon team with engineers, product leads, and architects across the MACH Alliance ecosystem.
