
SKT-OM is an advanced agentic Retrieval-Augmented Generation (RAG) system powered by a 13B parameter Large Language Model, built for intelligent and context-aware text generation. How it Works: Users input any natural language query or complex question. SKT-OM intelligently decomposes the query, retrieves relevant information using the SKT RAG framework, applies multi-step reasoning through LangGraph agents, and generates accurate, well-structured, and contextually rich responses. Key Capabilities: • Multi-agent LangGraph workflow (query analysis, decomposition, retrieval, reranking, reasoning, and final synthesis) • Advanced SKT RAG pipeline with query rewriting and contextual compression • Stateful conversations with persistent memory across interactions • Tool calling support for dynamic capabilities • Significantly reduced hallucinations and strong multi-hop reasoning ---Technical Architecture:--- •13B LLM trained and optimized on AMD Developer Cloud GPUs using $100 developer credits • Powered by ROCm 7.0 and AMD GPU And TEAM OM • Efficient inference with full support for agentic flows Impact: • SKT-OM demonstrates production-grade agentic RAG capabilities — capable of handling complex, knowledge-intensive queries with high accuracy and logical reasoning. It showcases the full potential of AMD hardware combined with open-source tools like SKT RAG and LangGraph for building next-generation AI agents. Fully developed during the AMD Developer Hackathon 2026 with regular build-in-public updates.
10 May 2026

SKT OMNI-ARC V49 is a production-oriented autonomous settlement platform that transforms manual receipt processing into a fully agentic workflow. The system ingests a physical receipt or invoice (via vision models), runs a 12-source hybrid AI pipeline (9 Gemini models + MoonDream local vision + AI/ML API + Featherless), performs logical validation, RBI-aligned compliance checks, fraud detection through AI memory, and executes instant USDC settlement using Circle Programmable Wallets on Arc Testnet. Key Capabilities: • Sub-second end-to-end processing in optimistic paths (demo shows <1 second from receipt to confirmed transaction) • Smart Document Detection (AI + OpenCV) to filter non-receipt images • Voice authorization using AssemblyAI for human-in-loop safety • USDC per-inference metering — true agentic economy where the AI pays for its own compute • Developer-controlled wallets with Entity Secret for secure autonomous signing • Live dashboard with green status indicators and real-time transaction feed We targeted Indian SMEs and export corridors where traditional cross-border payments take 24-72 hours and incur high fees. By combining powerful multimodal AI with Circle’s programmable money infrastructure and Arc’s fast stablecoin-native L1, we created a practical demonstration of how autonomous agents can handle real financial workflows safely and efficiently. The project showcases the power of combining Google Gemini models for fast reasoning, local models like MoonDream for privacy-sensitive vision tasks, and Circle + Arc for the money movement layer. Every transaction is auditable on-chain with full AI reasoning logs. This submission represents our vision for the Agentic Economy — where AI agents don’t just chat, they perceive, reason, comply, and settle real value autonomously.
26 Apr 2026