Field notes
Production AI, written plainly.
Notes on the unglamorous parts of AI systems: data contracts, runtimes, review loops, compute limits, and the engineering habits that keep them useful after the demo.
The Agentic Trap
The fastest way to make production AI fail is to let execution paths disagree.
The Runtime Boundary Is the Product
Whether the surrounding runtime can make the model usable, observable, recoverable, and repeatable.
The Critic Needs a Contract
A useful AI reviewer needs versioned behavior, constrained output, evidence, and failure metadata.
From Human Taste to Training Data
Human review only scales when the system can learn how to filter.
Scarce Compute Is a Product Constraint
A single GPU changes what can run together, what must wait, and what the system must admit.
Running Qwen3.6 27B INT4 on a 5090
A field note on local VLM critic runtime, quantization, startup, and structured output.
Benchmark Before You Optimize
AI generation speed becomes tractable when the team measures workflow phases.
Prompt Tokens Are Product Budget
Prompt length is the budget available for product decisions and constraints.
Manifests Beat Memory
Long-running AI work should depend on manifests the system can inspect, resume, and verify.
When Humans Can Move Files
The fix for ordinary storage failure is one clear source of truth.
The Trigger Token Is a Contract
A LoRA run is a contract between model, weights, trigger token, runtime, and review expectations.
When a LoRA Looks Worse
Treat a bad LoRA output like an investigation, not a verdict.
Do Not Train on Chaos
LoRA quality begins before training starts, with data prep the team can still explain.
The Remote Review Trap
Remote review should be a narrow surface over canonical data, not a second product.