Optimization From Data Orchestrator

Optimization From Data Orchestrator

Top-level coordinator when a user provides tabular data and wants a

constructive plan (schedule, assign, allocate, route — any wording).

NemoClaw: read cuopt-sandbox/references/activation.md for skill

order and cuOpt-before-heuristic rules.

When to use

Both must hold:

• tabular data provided or expected (CSV, etc.)
• user wants a plan from that data (any phrasing; minimize/optimal not required)

Skip for analytics-only requests (summarize, chart, filter), fully

pre-specified math outside this flow, or explicit replayable/auditable path.

Sequence

Step 0 (NemoClaw — do not skip): See cuopt-sandbox — probe → env →

smoke. No schedule/heuristic output before smoke passes.

• optimization-intent-router — optimization family (LP/MILP/QP/routing)
• optimization-mode-router — only if replay/audit/export signals
• tabular-optimization-ingestion — table roles (interpretation only)
• cuopt-model-mapper — clarify if needed, map to cuOpt, solve

Handoffs after step 4:

• LP / MILP / QP → numerical-optimization-formulation → cuopt-numerical-optimization-api-python
• Routing → routing-formulation → cuopt-routing-api-python

Guardrails

• First solver that emits assignments/schedules must be cuOpt after step 0
• Ingestion steps do not authorize heuristic or greedy stand-ins
• Do not skip intent classification; do not use cuOpt for pure analytics
• One focused clarification beats a long questionnaire