alexander@aros:~$ ls blog/ --sort date█
modifiedtagname
Nov 9, 2026
AI / ML
LLM-as-Judge: Calibrate It or Don't Ship It
A judge you have not calibrated is a random number generator that bills you. The biases, the rubric engineering, and the agreement math. Nov 9, 2026 AI / ML Prompt Injection Is a Design Constraint, Not a Bug
The model has one channel for instructions and data, and its usefulness comes from the same place its vulnerability does. Design around it. Nov 9, 2026 AI / ML Structured Outputs: The Cheapest Reliability Upgrade in the Stack
The gap between "usually returns JSON" and "is JSON by construction" holds a fleet of production failures. Closing it is nearly free. Nov 2, 2026 AI / ML What Is MCP? A Wire-Level Answer
Skip the USB analogy. The Model Context Protocol is a handful of JSON-RPC messages, and once you have read them you know when you need it and when you do not. Nov 2, 2026 AI / ML Agent Memory Is a Write Policy, Not a Database
Memory is four decisions: what to write, how to store it, what to read back, and when to forget. Only one is infrastructure, and it is the easy one. Nov 2, 2026 Data Data Leakage: How Models Cheat on Their Own Exams
The oldest failure in applied ML moved into your eval set, your fine-tune, and your RAG benchmark. The model has seen the answer key. Oct 25, 2026 AI / ML Price the Task, Not the Token
Dollars per million tokens is the sticker price. The unit your business actually buys is a finished task, and at task granularity the numbers get absurd. Oct 25, 2026 Failure Modes RAG Failure Modes: The Retriever Did It
Failure Modes #2. When a RAG answer is wrong, everyone blames the model. Six ways the failure happened earlier, and what catches each one. Oct 10, 2026 AI / ML Prompt Caching Is Free Money
Same model, same outputs, up to 90 percent off the repeated part of your bill. Why most teams still pay full price. Oct 10, 2026 AI / ML Most LLM Guardrails Don't Survive Production
They demo clean and die quietly: by latency, by false positives, or by a bypass nobody wrote down. What the surviving ones have in common. Oct 7, 2026 Failure Modes Your LLM System Is Degrading Silently. Here Is How to Catch It.
Failure Modes #1. No deploy, no alert, no stack trace. Just quality sliding a little every week while every dashboard stays green. Oct 7, 2026 Data Chunking Strategies, Measured: 800 Questions, Three Splitters, One Surprise
Fixed vs sentence-aware vs recursive chunking against real questions. Structure was free. Overlap was not what it claimed to be. Oct 7, 2026 AI / ML MCP vs. RAG Is the Wrong Question
One is a protocol. One is a pattern. Comparing them is like comparing USB to a filing system, and the real decision hiding underneath deserves a better frame. Oct 7, 2026 AI / ML Context Engineering Is Not Prompt Engineering with a Better Name
Prompt wording is the smallest lever in the system. What the model sees, in what order, under what budget: that is the job now, and it has rules worth learning. Oct 6, 2026 AI / ML Your LLM Evals Are Lying to You
Your eval suite passes. Your users still get garbage. The metrics that mean something, the framework landscape without the marketing, and the build order that closes the gap. Oct 5, 2026 AI / ML RAG vs. CAG vs. KAG: Pick the Boring One
RAG, CAG, and KAG compared by someone who runs them in production: what each actually is, what it costs, and a decision framework that ignores the acronyms. Feb 14, 2026 AI Strategy The $400B Lie: Why Most Enterprise AI Projects Fail Before They Start
After leading AI transformations at multiple enterprises, I've found the same fatal mistakes every time. The problem isn't the technology; it's the question you're asking it to answer. Feb 7, 2026 Architecture Multi-Agent Systems Are the New Microservices. And Nobody's Readyupdated
Multi-agent architecture patterns from eighteen months in production: orchestration vs. choreography, agent communication, failure modes, and deployment, with diagrams. Jan 29, 2026 AI / ML I Replaced Our Entire Data Pipeline with an LLM. Here's What Broke.
A first-person post-mortem on replacing a traditional ETL pipeline with LLM-powered processing. What worked, what exploded in cost, what hallucinated, and the hybrid architecture we finally settled on. Jan 18, 2026 AI / ML RAG Is Dead. Long Live Agentic Retrieval.updated
RAG vs. agentic retrieval, from production: why vanilla RAG hit its ceiling, what agents that plan their own retrieval look like, and the hybrid architecture that beat it by 40%. Jan 6, 2026 Frontend The AI-Native Frontend: How ML Models Are Eating the UI Layer
ML models are no longer just backends; they're reshaping the UI layer itself. Predictive prefetching, adaptive layouts, and AI-generated components are here. Dec 22, 2025 Career 16 Years in Tech: The Only 5 Things That Actually Mattered
From junior dev in Detroit to CEO of Ryshe; 16 years of hard-won lessons distilled into the only 5 things that actually changed the trajectory of my career. Dec 10, 2025 AI / ML Fine-Tuning is a Trap. Here's What to Do Instead.updated
RAG vs. fine-tuning vs. prompt engineering, settled with real numbers: 8 experiments, $40K in GPU costs, one honest success. A decision framework from production.
A judge you have not calibrated is a random number generator that bills you. The biases, the rubric engineering, and the agreement math. Nov 9, 2026 AI / ML Prompt Injection Is a Design Constraint, Not a Bug
The model has one channel for instructions and data, and its usefulness comes from the same place its vulnerability does. Design around it. Nov 9, 2026 AI / ML Structured Outputs: The Cheapest Reliability Upgrade in the Stack
The gap between "usually returns JSON" and "is JSON by construction" holds a fleet of production failures. Closing it is nearly free. Nov 2, 2026 AI / ML What Is MCP? A Wire-Level Answer
Skip the USB analogy. The Model Context Protocol is a handful of JSON-RPC messages, and once you have read them you know when you need it and when you do not. Nov 2, 2026 AI / ML Agent Memory Is a Write Policy, Not a Database
Memory is four decisions: what to write, how to store it, what to read back, and when to forget. Only one is infrastructure, and it is the easy one. Nov 2, 2026 Data Data Leakage: How Models Cheat on Their Own Exams
The oldest failure in applied ML moved into your eval set, your fine-tune, and your RAG benchmark. The model has seen the answer key. Oct 25, 2026 AI / ML Price the Task, Not the Token
Dollars per million tokens is the sticker price. The unit your business actually buys is a finished task, and at task granularity the numbers get absurd. Oct 25, 2026 Failure Modes RAG Failure Modes: The Retriever Did It
Failure Modes #2. When a RAG answer is wrong, everyone blames the model. Six ways the failure happened earlier, and what catches each one. Oct 10, 2026 AI / ML Prompt Caching Is Free Money
Same model, same outputs, up to 90 percent off the repeated part of your bill. Why most teams still pay full price. Oct 10, 2026 AI / ML Most LLM Guardrails Don't Survive Production
They demo clean and die quietly: by latency, by false positives, or by a bypass nobody wrote down. What the surviving ones have in common. Oct 7, 2026 Failure Modes Your LLM System Is Degrading Silently. Here Is How to Catch It.
Failure Modes #1. No deploy, no alert, no stack trace. Just quality sliding a little every week while every dashboard stays green. Oct 7, 2026 Data Chunking Strategies, Measured: 800 Questions, Three Splitters, One Surprise
Fixed vs sentence-aware vs recursive chunking against real questions. Structure was free. Overlap was not what it claimed to be. Oct 7, 2026 AI / ML MCP vs. RAG Is the Wrong Question
One is a protocol. One is a pattern. Comparing them is like comparing USB to a filing system, and the real decision hiding underneath deserves a better frame. Oct 7, 2026 AI / ML Context Engineering Is Not Prompt Engineering with a Better Name
Prompt wording is the smallest lever in the system. What the model sees, in what order, under what budget: that is the job now, and it has rules worth learning. Oct 6, 2026 AI / ML Your LLM Evals Are Lying to You
Your eval suite passes. Your users still get garbage. The metrics that mean something, the framework landscape without the marketing, and the build order that closes the gap. Oct 5, 2026 AI / ML RAG vs. CAG vs. KAG: Pick the Boring One
RAG, CAG, and KAG compared by someone who runs them in production: what each actually is, what it costs, and a decision framework that ignores the acronyms. Feb 14, 2026 AI Strategy The $400B Lie: Why Most Enterprise AI Projects Fail Before They Start
After leading AI transformations at multiple enterprises, I've found the same fatal mistakes every time. The problem isn't the technology; it's the question you're asking it to answer. Feb 7, 2026 Architecture Multi-Agent Systems Are the New Microservices. And Nobody's Readyupdated
Multi-agent architecture patterns from eighteen months in production: orchestration vs. choreography, agent communication, failure modes, and deployment, with diagrams. Jan 29, 2026 AI / ML I Replaced Our Entire Data Pipeline with an LLM. Here's What Broke.
A first-person post-mortem on replacing a traditional ETL pipeline with LLM-powered processing. What worked, what exploded in cost, what hallucinated, and the hybrid architecture we finally settled on. Jan 18, 2026 AI / ML RAG Is Dead. Long Live Agentic Retrieval.updated
RAG vs. agentic retrieval, from production: why vanilla RAG hit its ceiling, what agents that plan their own retrieval look like, and the hybrid architecture that beat it by 40%. Jan 6, 2026 Frontend The AI-Native Frontend: How ML Models Are Eating the UI Layer
ML models are no longer just backends; they're reshaping the UI layer itself. Predictive prefetching, adaptive layouts, and AI-generated components are here. Dec 22, 2025 Career 16 Years in Tech: The Only 5 Things That Actually Mattered
From junior dev in Detroit to CEO of Ryshe; 16 years of hard-won lessons distilled into the only 5 things that actually changed the trajectory of my career. Dec 10, 2025 AI / ML Fine-Tuning is a Trap. Here's What to Do Instead.updated
RAG vs. fine-tuning vs. prompt engineering, settled with real numbers: 8 experiments, $40K in GPU costs, one honest success. A decision framework from production.
j / k to move · enter to open · ⌘K for everything
Subscribe on Substack →
One honest AI essay every week. No hype, no vendor pitches.