Insights
Field notes from production AI
We're writing up what we learn shipping agents, training models, and running the infrastructure underneath — the specifics, not the press releases. First pieces are on the way.
What we're writing next
Why most agent demos never reach production
Evaluation, guardrails, and the failure modes nobody plans for.
Fine-tuning vs. RAG: choosing without the hype
A cost, latency, and accuracy comparison on real workloads.
Getting more out of one GPU cluster
Partitioning, scheduling, and the economics of shared inference.
Annotation quality is a model problem
Inter-annotator agreement, QA passes, and data lineage in practice.