·
Researchers & Thinkers
Athos Georgiou
Founder & CEO |NCA
Applied AI researcher publishing on LLM inference optimization on AMD GPUs and multimodal document retrieval. Builds production RAG and agent systems.
Hugging Face
AI Models
3 models on Hugging Face · 423 downloads
Research Papers
4 papersarXiv
Architecture-Aware LLM Inference Optimization on AMD Instinct GPUs: A Comprehensive Benchmark and Deployment Study
2603.10031·2026-02-27
arXiv
Spatially-Grounded Document Retrieval via Patch-to-Region Relevance Propagation
2512.02660·2025-12-02
DOI
Improving Performance of Taranta: Analysis of Memory Requests and Implementation of the Solution
10.18429/jacow-icalepcs2023-tupdp044·2023-12-16
DOI
Front-End Monitor and Control Web Application for Large Telescope Infrastructures: A Comparative Analysis
10.18429/jacow-icalepcs2023-tumbcmo09·2023-11-27
Biography
Deep Research
Athos Georgiou is an applied AI researcher and the Founder & CEO of NCA, specializing in vision-language retrieval, LLM inference optimization, and multimodal document understanding. He created Snappy (83 stars), a vision-language retrieval system that unifies ColPali late-interaction with structured OCR via patch-to-region relevance propagation — the subject of both an arXiv paper and an open-source release. His benchmark study on architecture-aware LLM inference for AMD Instinct MI325X GPUs spans models from 235B to 1 trillion parameters. Georgiou trains and publishes his own ColQwen retrieval models on Hugging Face, and actively contributes to the Claude Code ecosystem with plugins like kimchi-cult and agent-debate.
Vision-Language RetrievalColPali / ColQwen ModelsLLM Inference OptimizationMultimodal Document UnderstandingGraph RAGGPU Benchmarking (AMD Instinct)Open-Source AI ToolsResponsible AIClaude Code PluginsDocument OCR Processing
Timeline
10 Research3 HF8 GitHub21 total
2026
2026-03-15Research
Published blog post 'Diminishing Returns and the Art of Knowing When to Stop' on ColQwen3.5 optimization plateaus2026-02-27Research
Published arXiv paper 'Architecture-Aware LLM Inference Optimization on AMD Instinct GPUs'2026-02-15GitHub
Created larvling2025
2025-12-02Research
Published arXiv paper 'Spatially-Grounded Document Retrieval via Patch-to-Region Relevance Propagation'2025-10Research
Released Snappy, a vision-language retrieval system unifying ColPali with structured OCR (83 stars)2025-09Research
Published 'The Most Beautiful RAG' — end-to-end vision RAG with ColPali, Qdrant, and MinIO2025-07-28GitHub
Created Snappy: Vision-language retrieval system unifying ColPali with structured OCR2025-05-24GitHub
Created little-scripts: A monorepo of utility scripts and tools2025-03-17GitHub
Created nlp-graphrag-with-qdrant-and-neo4j2025-03-03GitHub
Created Qdrant-Neo4j-Ollama-Graph-Rag2025-02-27GitHub
Created colpali1.3-colqwen2.5-v0.1Key Contributions
Snappy
Vision-language retrieval system unifying ColPali late-interaction with structured OCR via patch-to-region relevance propagation. 83 stars on GitHub.
Architecture-Aware LLM Inference Optimization on AMD Instinct GPUs
Comprehensive benchmark spanning 235B to 1T parameter models, co-optimizing model architecture, hardware, and serving-system configuration on MI325X GPUs
ColQwen3.5-4.5B (v1–v3)
Series of visual document retrieval models trained and published on Hugging Face, iterating on ColQwen architecture for multimodal retrieval
Qdrant-Neo4j-Ollama Graph RAG
Graph RAG implementation combining Qdrant vector search with Neo4j knowledge graphs and Ollama for local LLM inference. 58 stars.
Titanium
Full-stack AI application framework built with Next.js, TypeScript, and LangChain. 47 stars on GitHub.
kimchi-cult
Claude Code plugin marketplace for discovering and sharing AI development tools
Raising Artificial Intelligence
Co-authored framework for responsible AI development, addressing ethics, regulation, and governance through a parent-child relationship metaphor
CopperAI
Hands-free voice-to-voice interaction system with LLMs, enabling real-time spoken dialogue. 29 stars on GitHub.
Notable Quotes
“
AI development much like parenting, comes with rights and responsibilities. Creators, users, and regulators have distinct rights and responsibilities.
“
It's a powerful statistical model capable of generating text, but it lacks intelligence, creativity, and consciousness. Achieving AGI will likely require a combination of several models and technologies, not just GenAI.
“
Helping organisations navigate AI adoption through applied research, performance engineering, and production-grade implementation.
12 sources(click to expand)
Athos Georgiou - AI Research & ConsultingAthos G - Founder & CEO @ NCA | LinkedInathrael-soju on Hugging Face (3 models, 1 dataset)athrael-soju on GitHub (25+ repos, Snappy 83★)Athrael.net — BlogArchitecture-Aware LLM Inference Optimization on AMD Instinct GPUs (arXiv)Spatially-Grounded Document Retrieval via Patch-to-Region Relevance Propagation (arXiv)Raising Artificial Intelligence by Olga Topchaya and Athos GeorgiouNCA LTD — Companies House SC795284Athos Georgiou's software portfolio | DevpostAthos Georgiou on INSPIRE-HEPORCID 0009-0004-3081-2883
Research generated March 18, 2026
Researchers & Thinkers/Athos Georgiou
All Profiles