Idea
A parameter-efficient fine-tuning platform adapting large language models for predictive process monitoring to improve accuracy and speed for enterprises.
Research Paper
Core Innovation
This paper introduces a method to adapt pretrained large language models directly to process data without converting event logs into natural language. It uses parameter-efficient fine-tuning to reduce computational costs while improving predictive accuracy. The approach outperforms traditional RNNs and narrative-style methods, especially in multi-task scenarios, with faster convergence and less hyperparameter tuning.
Market Size (TAM)
$2–10B TAM, $1–2B SAM; assumption: growing demand for AI-driven process optimization and predictive monitoring in enterprises.
Potential Customers & Pain Points
- Enterprises Using Process Mining Tools Needing More Accurate Predictions
- Software Vendors Offering Predictive Process Monitoring Solutions
- Data Scientists Struggling With High Computational Costs For Model Training
Business Model
Subscription-based SaaS platform offering API access and enterprise licenses for predictive process monitoring enhancements.
Competitive Landscape
- Celonis
- UiPath
- Automation Anywhere
Implementation Challenges
- Integration with existing enterprise systems
- Data privacy and security concerns
- Adoption resistance due to new AI methods
Validation Strategy
- Pilot integration with select enterprise process mining platforms
- Benchmark predictive accuracy against existing RNN and narrative methods
- Collect user feedback on ease of integration and tuning requirements
Research Paper Overview
Domain Adaptation of LLMs for Process Data
Summary
This study explores direct adaptation of pretrained Large Language Models to process data without converting event logs into natural language, focusing on parameter-efficient fine-tuning to reduce computational costs. Experiments in Predictive Process Monitoring show improved predictive accuracy over RNNs and narrative-style methods, especially in multi-task settings, with faster convergence and less hyperparameter tuning required.