← All writing

Follow a thread.

Browse the questions, methods, and product decisions behind the notes.

a2a

A2A Interoperability Still Needs a Product Contract

agents

Agent Memory Needs an Expiry Date Agent Observability Without Logging Everything Agent Reliability: One Success Is Not Enough Durable Agents Need More Than Checkpoints Human–Agent Handoffs Need Evidence and Authority MCP Connects Tools. It Does Not Authorize Every Action Prompt Injection Is a Data-to-Action Problem Small Models Need a Fair Job Description

ai product

An AI Release Gate Should Be a Decision Record Cost Per Useful Outcome Beats Cost Per Token Human–Agent Handoffs Need Evidence and Authority

architecture

When More Agents Make a System Worse

bibliometrics

Bibliometrics for Beginners: Reading Research Signals Carefully

context engineering

Long Context Is Not a Retrieval Strategy

cost

Cost Per Useful Outcome Beats Cost Per Token Test-Time Compute Is a Budget, Not a Magic Switch

data governance

Agent Memory Needs an Expiry Date

data quality

Data Contracts: The Less Glamorous Side of Reliable ML

delivery

Durable Agents Need More Than Checkpoints

documents

Document RAG Has to See the Page

engineering

Build Versus Buy Is an Ownership Decision From Prototype to Product: The Work Around the Demo

evaluation

Agent Reliability: One Success Is Not Enough An AI Release Gate Should Be a Decision Record Cost Per Useful Outcome Beats Cost Per Token Your LLM Judge Needs an Evaluation Too Hybrid Retrieval: Measure the Candidate Set First Test-Time Compute Is a Budget, Not a Magic Switch When More Agents Make a System Worse Zero-Shot Forecasting Is a Starting Point, Not a Waiver Treat Prompt Changes Like Software Releases When an AI Product Should Say “I Don't Know”How to Actually Evaluate an LLM Application Why RAG Isn't a Silver Bullet Notes from Building Multi-Agent Eval Harnesses The Quiet Cost of Average Accuracy

evidence

AI Horizon Scanning Needs a Counterevidence Lane Evidence Graphs Before GraphRAG

experimentation

Product Experiments When You Don't Have Huge Traffic

forecasting

Zero-Shot Forecasting Is a Starting Point, Not a Waiver

foresight

AI Horizon Scanning Needs a Counterevidence Lane Building a Trend Radar with LLMs (Without Hallucinating the Future)What a Trend Radar Actually Does in a Boardroom Bibliometrics for Beginners: Reading Research Signals Carefully Strategic Foresight vs Forecasting: They're Not the Same The Patent Signal: What Filings Quietly Reveal Designing for the 10-Year Question

foundation models

Zero-Shot Forecasting Is a Starting Point, Not a Waiver

genai

AI Horizon Scanning Needs a Counterevidence Lane

graphrag

Evidence Graphs Before GraphRAG

human-ai interaction

Human–Agent Handoffs Need Evidence and Authority Human Review Is a Workflow, Not a Checkbox When an AI Product Should Say “I Don't Know”From Black Box to Glass Box: Why Multi-Agent Systems Need to Show Their Work

interoperability

A2A Interoperability Still Needs a Product Contract

lessons

What Air Cargo Forecasting Taught Me About Real-World ML

llm judges

Your LLM Judge Needs an Evaluation Too

llms

Before You Add an LLM, Build the Simplest Useful Baseline Designing a Latency Budget for AI Features RAG, Fine-Tuning, or Neither? Start with the Failure Treat Prompt Changes Like Software Releases Building a Trend Radar with LLMs (Without Hallucinating the Future)How to Actually Evaluate an LLM Application

long context

Long Context Is Not a Retrieval Strategy

mcp

MCP Connects Tools. It Does Not Authorize Every Action

memory

Agent Memory Needs an Expiry Date

methods

Strategic Foresight vs Forecasting: They're Not the Same

metrics

ML Monitoring After Launch: From Dashboards to Decisions Product Metrics That Change a Decision The Quiet Cost of Average Accuracy

mlops

Data Contracts: The Less Glamorous Side of Reliable ML ML Monitoring After Launch: From Dashboards to Decisions Treat Prompt Changes Like Software Releases Production Time Series in 2026: The Boring Tools That Win

model routing

Small Models Need a Fair Job Description

multi-agent

From Black Box to Glass Box: Why Multi-Agent Systems Need to Show Their Work Notes from Building Multi-Agent Eval Harnesses

multi-agent systems

A2A Interoperability Still Needs a Product Contract When More Agents Make a System Worse

multimodal ai

Document RAG Has to See the Page

mvp

An MVP Should Complete One Workflow

observability

Agent Observability Without Logging Everything

patents

The Patent Signal: What Filings Quietly Reveal

performance

Designing a Latency Budget for AI Features

prioritization

Turning Product Feedback into Decisions, Not Just Tickets

privacy

Agent Observability Without Logging Everything

product delivery

Human Review Is a Workflow, Not a Checkbox An MVP Should Complete One Workflow From Prototype to Product: The Work Around the Demo

product discovery

Before You Add an LLM, Build the Simplest Useful Baseline Product Discovery Before the Backlog Gets Expensive Turning Product Feedback into Decisions, Not Just Tickets Product Metrics That Change a Decision

product strategy

Build Versus Buy Is an Ownership Decision

rag

Document RAG Has to See the Page Hybrid Retrieval: Measure the Candidate Set First Long Context Is Not a Retrieval Strategy RAG, Fine-Tuning, or Neither? Start with the Failure Why RAG Isn't a Silver Bullet

reasoning

Test-Time Compute Is a Budget, Not a Magic Switch

release management

An AI Release Gate Should Be a Decision Record

reliability

Agent Reliability: One Success Is Not Enough Durable Agents Need More Than Checkpoints Your LLM Judge Needs an Evaluation Too

retrieval

Evidence Graphs Before GraphRAG Hybrid Retrieval: Measure the Candidate Set First

security

MCP Connects Tools. It Does Not Authorize Every Action Prompt Injection Is a Data-to-Action Problem

small models

Small Models Need a Fair Job Description

strategy

What a Trend Radar Actually Does in a Boardroom Designing for the 10-Year Question

time-series

What Air Cargo Forecasting Taught Me About Real-World ML Production Time Series in 2026: The Boring Tools That Win

tool use

Prompt Injection Is a Data-to-Action Problem

user research

Product Discovery Before the Backlog Gets Expensive Product Experiments When You Don't Have Huge Traffic