Author

Owen Park

AI security researcher

Covers AI application security, model risk, and agentic system abuse.

ai-securityresearchapplication-security

Coverage by Owen Park (28)

Developer workstation showing IDE and terminal agents as parts of a 2026 AI coding stack

RTX PRO 6000 Blackwell 96GB: The Local LLM Workstation Buyer's Guide

A workload-first reasoning guide to the 96 GB single-address-space card: what genuinely fits, where MoE models break it, and when a cluster or rental wins.

Infographic showing DeepSeek V4.1 Flash served across three DGX Spark nodes with key throughput figures

Nvidia Dgx Spark Research

Multi-Node DGX Spark: Tensor-Parallel Sizing for 2x and 4x Clusters

A 135 GB model against a 128 GB box forces a cluster. Here is the decision order, the parallelism maths, and what the interconnect really costs per token.

Illustration comparing China's Kimi K3 open-weight AI model against Western coding assistants Claude and Codex.

Agent Harness AI Security

Agent Harness Architecture: How Coding Agents Actually Execute Tools

A mechanism-first breakdown of the runtime that wraps a model: the loop, tool schemas, sandboxing, context strategy, permissions, and injection defence.

Conceptual 3D illustration of sparse Mixture-of-Experts architecture routing in Laguna S 2.1.

Vllm Vs Sglang Research

SGLang vs vLLM vs TensorRT-LLM: Choosing a Production Inference Engine

Paged KV, radix-tree prefix sharing and ahead-of-time compilation all pull in different directions. Work out which one your own workload should pay for.

Open-weight AI model fit guide for a 24GB Apple Silicon Mac in 2026

Kv Cache Vram Research

KV Cache Explained: Why Long Context Costs More VRAM Than Your Weights

A model that fits can still die on one long prompt. Here is the arithmetic behind the second allocation that grows with context, and how to shrink it.

Protocol schematic poster for a 128GB Apple Silicon local AI stack led by Qwen3.6

Gpu Vram Requirements Research

Will It Fit? VRAM Math for Open-Weight Models Across Every Quantization Format

Weights are only one of four memory buckets, and past roughly 115k tokens the KV cache costs more than the model does. Here is the arithmetic in full.

Infographic mapping one to four NVIDIA DGX Sparks to the recommended local model at each node count: Qwen3.8 Flash Next, DeepSeek V4 Flash, GLM-5.3 Flash, DeepSeek V4.1 Flash.

Inference Engineering Research

Inference Engineering: A Practical Primer for Serving Your Own Models

Prefill is compute-bound, decode is bandwidth-bound, and KV cache is what really caps concurrency. The metrics, memory maths and tradeoffs behind serving.

Illustration of a DDR5 RAM module depicted as a soaring price peak towering over small PCs and a server rack, symbolizing surging memory prices.

Nvfp4 Quantization Research

NVFP4 vs FP8 vs MXFP4 vs EXL3: The Quantization Format Guide for Local LLMs

The same 320B model is 328 GB in FP8 and near 135 GB at 4 bits. Bytes-per-parameter math, per-format tradeoffs, and how to size a box that actually fits.

Desks covered

AI Security Research

Latest

Sep 12, 2026, 12:37 PM EDT · Research · Mlx12 minMLX vs llama.cpp on Apple Silicon: Which Local LLM Runtime to Actually Use

One is a portable inference engine, the other an array framework that also trains — and that single difference settles almost every practical question.

Sep 12, 2026, 12:37 PM EDT · Operations · Llm Evaluation14 minHow to Build an LLM Evaluation Harness That Catches Real Regressions

Public benchmarks only shortlist a model; your own cases decide whether to ship it. Here is the dataset, the scorers and the CI gates that catch real drift.

Sep 12, 2026, 12:37 PM EDT · Research · Prompt Caching11 minPrompt Caching Explained: Cutting Agent Costs Without Touching the Model

Your agent re-sends the same long prefix on every step and pays to recompute it. Here is the KV mechanism, the ordering rule it forces, and the break-even math.

Sep 12, 2026, 12:37 PM EDT · Research · Speculative Decoding11 minSpeculative Decoding and Multi-Token Prediction: How LLMs Generate Faster

Decode is memory-bound, so checking several tokens in one pass is nearly free. Here is the mechanism, the arithmetic behind it, and why quality never moves.