// 01 audit
Model audit
Baseline quality, cost, and risk of your current stack — golden set sample, latency p95, and failure modes written down.
eval_baseline cost_map risk_register
// pipeline modules
// services · model_stackSelection. Fine-tune. Eval. Serve.
Modules that ship measured quality — not “we fine-tuned something.” Local GPU capacity when corpora cannot leave; cloud models and multi-model routes when APIs win on quality and cost.
// 02 finetune
Fine-tune program
Dataset hygiene, training loops, and acceptance tests. Loss curves tracked; overfit caught before it hits prod.
val_loss↓ golden_set local_gpu?
// 03 serve
Private serving
On-prem or VPC inference with monitoring. When data cannot leave, we serve on local GPU with ops you can staff.
p95_latency healthchecks rollbacks
// 04 route
Router design
Multi-model routing for cost vs quality — frontier APIs, smaller open weights, or specialist heads behind one interface.
multi_model cost_cap fallback
acceptance
metrics first
regression
gate:on
privacy
local_gpu
handoff
your team owns