RAG Is Not a Silver Bullet
Retrieval-augmented generation is the default architecture for knowledge-grounded AI. But it fails in ways most teams don't anticipate until production.
AI
Notes from building with large language models: evaluation pipelines, tool use in agents, retrieval, prompt design, and the quiet ways AI products fail.
Building Reliable Tool-Use in LLM Agents
Tool calls are the most fragile part of agentic LLM workflows. Here are patterns that make them predictable: schema design, validation, and retries.
The Hidden Cost of AI Wrapper Products
Thin wrappers around LLM APIs aren't defensible products. Here's why they fail and what to build instead.
Prompt Engineering Is Software Engineering
Prompts are production artifacts. They deserve version control, testing, review processes, and deployment strategies, just like any other code you ship.
When to Eval: Building Evaluation Pipelines for LLM Features
Most teams ship LLM features without systematic evaluation. Here's when evals matter, how to keep them lightweight, and how to catch regressions early.
AI Products Don't Fail Loudly, They Fail Quietly
Why debugging AI systems is less about fixing errors and more about auditing decision-making.
Building Agentic AI for Personal Finance: Lessons Learned
What I learned designing an agentic AI system to prioritize and forecast spending, with guardrails that keep it explainable, auditable, and context-aware.
Single Prompt vs Step-Based LLM Workflows: A Design Tradeoff
Exploring the tradeoff between single-prompt convenience and step-based control when designing LLM-powered features.
My 2025 Wrapped: What I Built, Shipped, and Learned in AI
A look back at the AI side projects I built in 2025 and the technologies that shaped how I think about shipping AI into production.