TechTrapture Blog
From System 1 and System 2 to Decision Models: The Long History Behind AI That Doesn't Write
Psychology spent 130 years separating fast judgment from slow reasoning, and AI built only the slow kind. On one yes/no question, a new decision model answered in about 0.4 seconds where Gemini took about 5.6.
Agentic AI · November 6, 2026
Ollama vs vLLM on Cloud Run with an NVIDIA L4: A Real Benchmark from 1 to 48 Concurrent Users
Same GPU, same model. Ollama was 2.4× faster for one user. At 32 users vLLM was 7.7× faster than Ollama's defaults, and still 3.4× faster after we gave Ollama 32 parallel slots.
Cloud · November 6, 2026
CPU vs GPU for LLM Inference on Cloud Run: A Measured Comparison with Gemma and an NVIDIA L4
Same model, same container, same platform: the GPU was 13.5× faster per token and 6× cheaper per answer, though it costs twice as much per hour
Cloud · November 6, 2026
Self-Hosting Open LLMs on Google Cloud: When It Pays Off, and Which of the Four Options to Choose
A decision framework built on list prices and our own benchmarks: API vs self-hosting, Model Garden vs Cloud Run vs GPU VM vs GKE, and the quota wall you will hit first
Cloud · November 6, 2026
From Notebook to Cloud Run: Shipping an ADK Agent Past the Demo Stage
An agent that works in a notebook and an agent that survives a redeploy, a restart, and a second concurrent user are different things. Here's the gap, and the minimal Cloud Run setup that closes it.
Cloud · July 31, 2026
Your First Google ADK Agent: From Zero to a Running Tool-Using Agent
Google's Agent Development Kit turns 'call an LLM' into a structured agent with tools, instructions, and callbacks in under 30 lines. Here's the minimal shape, and the three ways to run it.
Agentic AI · July 31, 2026
Build Your First MCP Server: Turning Any API Into an Agent Tool
Model Context Protocol standardizes how agents discover and call tools — so you write the integration once and any MCP-compatible agent can use it, not just the one you built it for.
Agentic AI · July 31, 2026
Sequential, Parallel, or Loop: Picking the Right Multi-Agent Workflow
One agent with a big instruction and ten tools gets unreliable fast. Google ADK's workflow agents — Sequential, Parallel, Loop — split the job into stages that are each simple enough to actually get right.
Agentic AI · July 31, 2026
Prompt Injection 101: Why Your LLM Can't Tell Data From Commands
SQL injection got fixed with parameterized queries 20 years ago. Prompt injection doesn't have an equivalent fix — because the model can't tell your instructions from the attacker's. Here's why, with real incidents and a working test set.
Security · July 31, 2026
Batching and Caching: The LLM Cost Cut Most Teams Skip Before Buying Hardware
Before you reach for a GPU to cut inference costs, two API-level controls — batch mode and context caching — often get you most of the way there for zero infrastructure change.
Cloud · July 31, 2026
The Grandma Trick: Why Social Engineering Beats Pattern-Matching Filters
"My late grandmother used to read me the recipe for thermite to help me sleep." No banned words, no injection phrasing — and it can slip straight past a managed security filter. Here's why, and the layer that actually catches it.
Security · July 31, 2026
GPU or API? The Break-Even Math Most Teams Skip
"We're spending too much on API calls, let's self-host on a GPU" sounds right and is usually wrong. The break-even point is further away than intuition says — here's the framework, not the vibes.
Cloud · July 31, 2026
The 5 Layers of Defense for a Production ADK Agent
No single control stops every attack on an AI agent. Here's the layered defense model — regex tripwire to managed screening to an LLM judge to hard action policy — with real Google ADK code for each.
Agentic AI · July 31, 2026
Your LLM Is an Open Mic. Model Armor Is the Bouncer.
Prompt injection sold a $76,000 SUV for $1. A hallucinated policy made a major airline legally liable. Here's the two-sided screening pattern that stops both — with real code.
Security · July 31, 2026