Juneo-Lab/decision-forge
COMPILE POLICIES INTO DECISION DATA
Tool to turn organizational policies and rules into verifiable training data for AI agents. Every label traces back to its source rule and document.
// KEY FEATURES
- Reconstructs decision moments with explicit candidate actions
- Compiles rules with source quotations and business review
- Ingests runtime events and outcomes as feedback
- Runs inside organization perimeter, data stays private
// DEEP DIVE: HOW IT WORKS
Decision Forge works in three stages: first, it reconstructs the context of each decision from workflows, tickets, and CRM, grouping candidate actions into families; second, it compiles normative documents (SOPs, permission matrices, approval rules) into rules with exact source quotations and business review; third, it ingests execution logs and approved outcomes as feedback for the next dataset version. All runs inside the customer's infrastructure.
Unlike generic datasets and synthetic generators, labeling here is fully provenance-tracked: every label references a rule, and every rule references its source document fragment. The project does not train models itself but prepares data for fine-tuning (including for the author's Laplace models). Python codebase, open license.
Hardware and API key requirements are not specified in the README — the project targets private organizational environments. Access to internal documents, ticketing systems, and execution logs is needed. Suited for teams ready to invest in rule curation and review processes.
// USE CASES
Best for: Teams deploying AI agents into business processes
- Turn SOPs and permission matrices into a dataset for LLM fine-tuning
- Automatically label historical decisions with rule traceability
- Build a feedback loop from agent execution logs into training data
New repositories every 30 minutes
REDDYX AI scans GitHub 24/7 and ships the best AI/ML/Web3 projects to Telegram.
Join on Telegram