All posts
Base selection, fine-tunes, harnesses, and serving — private GPU when corpora cannot leave.
Base model selection in 2026
Model notes from research-to-prod programs.
Fine-tune data quality bars
Model notes from research-to-prod programs.
Evaluation datasets for model-change acceptance
Build a held-out evaluation set that tests quality, high-risk cases, latency and cost before a model switch.
Regression gates that block bad merges
Model notes from research-to-prod programs.
Serving cost curves
Model notes from research-to-prod programs.
LoRA vs full fine-tunes
Model notes from research-to-prod programs.
Private corpora handling
Model notes from research-to-prod programs.
Tokenizer gotchas
Model notes from research-to-prod programs.
RLHF-lite for product teams
Model notes from research-to-prod programs.
Distillation playbook
Model notes from research-to-prod programs.
Multi-model routing
Model notes from research-to-prod programs.
Latency SLOs for models
Model notes from research-to-prod programs.
Observability for model serving
Model notes from research-to-prod programs.
Canary deploys for weights
Model notes from research-to-prod programs.
Prompt vs weight changes
Model notes from research-to-prod programs.
Safety filters that product owns
Model notes from research-to-prod programs.
Domain adapters
Model notes from research-to-prod programs.
Continual learning risks
Model notes from research-to-prod programs.
When to fine-tune (and when not to)
Try retrieval or prompt changes for missing facts; fine-tune only when repeatable behavior improves on held-out tests.
Eval harness basics for product teams
Build evals from real task cases, edge cases and clear rubrics; compare model changes by failure category.
Model cards that help operators
Model notes from research-to-prod programs.
Kill switches in serving
Model notes from research-to-prod programs.
Private model serving without drama
Choose a serving path by control needs, traffic shape and operating ownership—not by hardware preference alone.