Automation library · Data collection

Tune ROV PPO reward weights with OpenAI and Google Sheets

Started by another system calling it, and works across Gmail and Google Sheets. It hands the judgement calls to OpenAI. 17 steps, 3 branches, 1 loop or waits: a working afternoon to understand properly.

MediumCalls a model: OpenAI

Download free
17 stepsMedium: a working afternoon
3Branches: places the flow splits
1Loop or waits
2Services it connects to

The flow

Starts it 1Model 3Logic 7Service 6

The steps in the order they run, from what starts it.

  1. 1Start Tuning RunWebhookStarts it
  2. 2OpenAI Chat ModelLm Chat Open AiModel
  3. 3Parse Reward ProposalOutput Parser StructuredModel
  4. 4Set Run ConfigSetLogic
  5. 5Propose Reward WeightsChain LlmModel
  6. 6Prepare Round RequestSetLogic
  7. 7Validate Weights and Update HistoryCodeLogic
  8. 8Start PPO Training RoundHttp RequestService
  9. 9More Rounds?IfLogic
  10. 10Wait Before PollingWaitLogic
  11. 11Send Final ReportGmailService
  12. 12Check Training StatusHttp RequestService
  13. 13Training Finished?IfLogic
  14. 14Fetch Round MetricsHttp RequestService
  15. 15Training Failed?IfLogic
  16. 16Log Round to SheetGoogle SheetsService
  17. 17Send Failure AlertGmailService

How it's built

Starts on
Another system calling it
Steps
17, with 3 branches, 1 loop or waits
Size
Medium: a working afternoon
Connects to
Gmail, Google Sheets
Model
OpenAI

Questions

What does it need to run?

An n8n account, plus your own Gmail and Google Sheets accounts, and a key for OpenAI. You connect your own accounts after importing it; none are included.

What starts it?

It starts on another system calling it.

Does it use AI?

Yes. It calls OpenAI. Model calls are billed by the provider, not by n8n.

How long does it take to set up?

Plan an afternoon. 17 steps is enough that you should read each one before running it.

How do I import it?

Copy the workflow, open a new workflow in n8n and paste. The steps appear on the canvas, ready for your own accounts.

Before you run it

Nothing here has been tested by us. It is somebody's working automation, read and described from its own steps. Treat the connections it wants as a list of accounts it will touch, and read the 17 steps before you point it at anything that matters.