Products

FineForge

Turn drift signal into a fine-tune or prompt/adapter job you can ship. FineForge owns eval gates, canary promote, rollback, and atomic prompt+adapter+route bundles.

When to use

DriftWatch opened an alert and you need a versioned fix
Cheap SLM needs distillation from a teacher
Tool schema broke and adapters + fixtures must ship together

Job kinds

Kind	Artifact	Gate
prompt_patch	prompt_version	Golden pass rate
adapter_fix	tool_schema_version + fixtures	Contract tests
teacher_critique	Repair suggestions → patch	Judge agreement
lora_train	owned model revision	Eval + cost gap accept
bundle	prompt + adapter + route rules	Atomic promote
eval_expand	New cases from incidents	Coverage↑

Promote path

1. Job from DriftWatch alert or schedule
2. Scrub dataset (PII block — F11)
3. Eval harness (CI gate — F06)
4. Shadow / canary 1–5% (F14)
5. Full promote + deploy marker (F15)
6. On miss: rollback to previous revision (F02)

Metrics: fineforge.jobs.*, canary score, eval pass rate
Register new model_id (*-vN) in the model registry before RouteIQ points traffic
Postmortem → eval case (F24) so the failure never returns silently

Operator failure modes

ID	Problem	Next action
F02	Bad prompt deploy	Instant previous version
F07	Noisy LLM-as-judge	Rubric + few-shot; freeze
F11	PII in train set	Block train until scrubbed
F12	Noisy prod traces	MinHash + score filter
F14	Canary unclear	Hold until delta > ε
F16	Partial bundle ship	Atomic multi-artifact only
F20	Distill quality gap	Accept gap for cost; document
F23	Optimize $ without quality	Quality-weighted cost objective

API

POST /v1/fineforge/jobs
POST /v1/fineforge/jobs/{id}/promote
POST /v1/fineforge/jobs/{id}/rollback

See API reference and Drift & retrain.

Fits the loop

FineForge closes observe → route → drift → retrain. Owned models (guide) are clients of Inferix — FineForge promotes revisions; RouteIQ selects them.

Source

https://github.com/AkshantVats/fineforge

Operator and agent playbooks.

Guides →