Sherriff

GIF
A visualization of the generated news reports
LLM Analysis Framework
Initial testing is promising!!!!
Results

Inspiration

We were interested in one primary question: "Can new reports help predict disruptions in supply chains?"

Companies like Palantir often have extremely large datasets or incredibly comprehensive internal data streams. However, in the real world, this is often not always accessible, and humans in decision-making roles must make the best out of sparse, proprietary dataset.

What it does

We structured our project around three core components:

Simulating Supply Chain Events
Using a probabilistic event generator, we created a synthetic history of supply chain disruptions. Each node in our simulated supply chain experienced varying levels of disruption, which compounded over time.
Generating News Reports
The simulation produced synthetic news articles reflecting the disruptions at different supply chain nodes. This mimicked real-world reporting on supply chain issues.
Risk Forecasting with AI
We experimented with both traditional neural networks and large language models (LLMs) to analyze the news reports and predict structured risk scores. Despite having only news data, our approach performed comparably to traditional methods.

How we built it

We built it using langchain, python, openai, flask, and lots of code.

What we learned

Throughout the process, we uncovered several key insights:

News-based forecasting is viable – Sentiment analysis and structured extraction from news articles can offer meaningful risk indicators.
Simulated data can be useful – While real-world datasets are often sparse or proprietary, generating simulated historical supply chain events allowed us to test our models effectively.
Traditional neural networks struggle – Standard machine learning approaches are ok at making accurate risk predictions, but LLM-based methods performed surprisingly well with minimal input.