ContractRiskAI by machine-intelligence-group

Inspiration

Australian Government contracts can increase significantly after they are signed. The Australian National Audit Office found $182.2 billion in contract amendments over ten years. We wanted to help procurement teams identify which contracts deserve closer attention without manually reviewing every document in the same depth.

What it does

ContractRiskAI automates the first stage of contract risk review. A user uploads a contract PDF, and the system:

Extracts important contract details
Estimates the chance of a 25% or greater value increase
Estimates possible additional financial exposure
Finds similar Australian Government contracts
Generates clear review advice and recommended actions It supports procurement professionals while keeping them responsible for the final decision.

How we built it

We collected and prepared 50,000 public contracts using the official AusTender API. Original contract details were used as model inputs, while later amendments were used to determine what happened afterwards.

We used:

40,000 contracts for training
5,000 contracts for model selection
5,000 later contracts for testing

We combined XGBoost and CatBoost for risk prediction and used another CatBoost model to estimate financial exposure. We also built a historical search system that retrieves similar contracts. These results are passed to Qwen3-14B Instruct, which converts the contract details, predictions and historical evidence into actionable review advice. The application uses React, FastAPI and Modal, with Qwen hosted privately on an A100 GPU.

Challenges we ran into

One major challenge was processing long and inconsistent contract PDFs without ignoring important information in later pages. We solved this by dividing documents into smaller sections and processing multiple sections in parallel. We also had to prevent future contract information from leaking into model training, connect multiple models into one reliable pipeline, and keep users informed while an analysis was running.

Accomplishments that we're proud of

Our risk model achieved a ROC-AUC of 0.888 on 5,000 later test contracts. The top 10% of contracts selected by our dollar model accounted for 87.3% of all additional value recorded across the test set.

We are also proud that ContractRiskAI is more than a prediction model. It is a complete working product that processes real PDFs, produces tested predictions, retrieves verifiable AusTender evidence and turns everything into practical guidance.

What we learned

We learned that language models are most useful when they are grounded in reliable data and specialised models.

Instead of asking an AI to invent a risk assessment, we provide Qwen with tested predictions, structured contract information and relevant historical cases. This produces advice that is more consistent, useful and accountable.

What's next for ContractRiskAI by machine-intelligence-group

ContractRiskAI is currently designed for Australian Government procurement. Our next step is to test it with more contract formats, improve the accuracy of financial exposure estimates, and conduct trials with procurement professionals.

The same data preparation and training pipeline could also be extended to other countries when suitable local contract records and outcomes are available.