Author
Sebastian Mondragon
CEO & Founder
AI systems architect and founder of Particula Tech, building production-ready AI solutions for enterprises worldwide.
AI System ArchitectureLarge Language ModelsRAG SystemsAI Agent DevelopmentEnterprise AI StrategyLLM Evaluation
About
Sebastian Mondragon is the founder and CEO of Particula Tech, an AI consulting and development company that has delivered AI solutions to clients worldwide since 2023.
With deep expertise in building production-ready AI systems, Sebastian specializes in architecting solutions that handle real-world scale and complexity. His approach combines rigorous engineering discipline with practical business understanding, focusing on AI implementations that actually ship to production rather than remaining proof-of-concepts.
Before founding Particula Tech, Sebastian worked extensively on enterprise software systems, developing a strong foundation in building reliable, scalable applications. This background shapes his philosophy that AI systems need the same engineering rigor as any other production software: proper architecture, comprehensive testing, monitoring, and maintenance plans.
Sebastian writes regularly about AI development best practices, from technical deep-dives on RAG systems and AI agents to strategic guidance on when and how businesses should adopt AI. His writing reflects hands-on experience solving real problems for clients across manufacturing, professional services, healthcare, and retail industries.
His research preprint, "Can You Trust Your Own Eval? A Framework and Controlled Validation for Measuring Exploitability and Contamination in LLM Evaluation Harnesses" (2026, DOI: 10.5281/zenodo.20736881), examines how contamination and exploitability undermine private LLM evaluation systems and proposes reproducible measurement protocols for sealed-harness design.
Articles by Sebastian · 216
All articles
AI SECURITY01
Confidential GPU Inference: The Real H100 Performance TaxPublished H100 confidential-mode overhead spans 0% to 70%. Here is which number describes your serving stack, and the GPU-count delta it implies.
AI FOR BUSINESS02
EU AI Act Article 50: Marking AI Output in ProductionArticle 50 applies 2 August 2026, with fines to €15M or 3% of turnover. The four duties, the 2 December 2026 transition, and a marking pipeline that ships.
AI DEVELOPMENT TOOLS03
Full Prompt Re-Processing: llama.cpp Cache Fixes That WorkOne rejected request wiped 15 context checkpoints and 2.19 GiB of KV state in a single step. The llama.cpp cache flags that actually exist today.
AI DEVELOPMENT TOOLS04
MCP Stateless Migration 2026: What Breaks in a Live ServerEleven breaking changes landed in the 2026-07-28 MCP revision. Sessions, initialize, and Last-Event-ID are gone. The migration order that keeps clients live.
LLMS & MODELS05
Self-Host Kimi K3: GPU Sizing for 1,561 GB of WeightsKimi K3's checkpoint is 1,560.9 GB, not the 594 GB or 1.4 TB every guide printed. 8x H200 cannot load it. The shard-sum command and the vLLM flags.
AI SECURITY06
vLLM Security Hardening: The 0.24.0 Floor for CVE-2026-22778CVE-2026-22778 scored 9.8 and shipped for 302 days. Patching to 0.14.1 does not close it: the real floor is 0.24.0, and your API key was bypassable.