NPUs quietly won the year
The loudest AI news is cloud. The quietest news is a Snapdragon laptop running an 8B model at 30 tokens per second. On-device inference just went from novelty to default.
FIELD NOTES
What we're reading. What we're shipping.
Weekly notes on building with AI, putting agents in products, AWS AI, security, and compliance — written for operators, not slide decks. Updated ~once a week.
Agents that ship — and agents that demo
The industry is drowning in agent demos. Here's the boring checklist that separates a terminal trick from something you can leave running overnight.
AWS AI without the bill shock
Bedrock, SageMaker, and friends are powerful. They're also easy to set on fire. How we size, cache, and kill runaway spend on real products.
RAG that doesn't lie to your users
Retrieval-augmented generation is still the default pattern for company knowledge. Most implementations hallucinate politely. Fixes that actually move the needle.
Prompt injection is not a meme
If an agent can read email and call tools, an attacker can try to become the prompt. Offense and defense we run on every Embed and Edge engagement.
Evals before vibes
Shipping on “it felt good in the playground” is how demos die in prod. A minimal eval harness for product teams who don't have an ML platform org.
Fine-tune or don't
Fine-tuning is romantic. Prompting + RAG + tools is usually cheaper. When we actually recommend a custom model — and when we don't.
AI compliance is an audit trail
Regulators and customers don't care that your model is clever. They care who approved the action, what data touched the prompt, and whether you can prove it later.
Machine learning that ships in weeks
Classical ML still wins for ranking, fraud, and forecasting. How we blend old-school models with LLM agents without turning the roadmap into a research lab.
Context isn't free
Million-token windows are a sales slide, not a strategy. What long context actually costs, where it rots, and when retrieval still wins.
Reasoning tokens are observability
Extended thinking gets treated as an expensive party trick. Read the traces and it becomes a debugger for prompts.
Small models, big deployment
Everyone reaches for the flagship. Haiku, Flash, and their peers do 80% of the work at 3% of the price. The trick is knowing which 80%.
Embeddings that don't drift
Your RAG worked in demo. Six months later, top-1 is garbage. What changed — and how to catch it before your users do.
The framework tax
Every quarter another agent framework launches, and two die. What we reach for instead — and when a framework is actually the right answer.