ContractScanner build note

The operating problem

Contracts and NDAs often need a fast first-pass review before a founder, freelancer, procurement lead, or legal professional decides which clauses deserve attention. Many online review tools require the user to upload sensitive text to a remote service before they can learn whether the document contains unusual obligations.

The product question was whether a useful risk-triage experience could run inside a supported browser, keep the contract text on the user's device, and still present structured issues, missing protections, and negotiation prompts in a form a human could review.

How we approached it

ContractScanner downloads an ONNX language model to a compatible WebGPU browser after the user chooses to load it. Contract text is then analysed locally. The interface converts model output into a score, risk sections, clause references, missing protections, and a plain-language verdict while preserving the raw output for inspection.

The tool is deliberately described as triage rather than legal advice. It helps a user identify questions and potentially one-sided terms; it cannot determine the commercial context, applicable law, negotiating position, or professional advice appropriate to a significant agreement.

What was built

• A capability check for WebGPU and available device memory before model loading.
• On-demand model download rather than loading the model with the landing page.
• Local browser inference so pasted contract text is not sent to a SupaMakers API.
• A structured review format covering risk severity, clause location, issue, recommended change, and missing protections.
• A synthetic sample agreement that allows visitors to inspect the workflow without using a real contract.
• Visible limitations and recovery states for unsupported browsers, model-loading failures, and analysis errors.

Constraints and trade-offs

• The first model load requires a network download and can be substantial; local analysis begins only after that model is available in the browser.
• WebGPU support, device memory, and browser behaviour vary, so the tool cannot offer the same experience on every device.
• Language-model output is probabilistic and can miss a clause, misread context, or suggest an inappropriate change.
• A risk score is a triage signal, not a legal conclusion. Significant agreements still require review by an appropriately qualified person.

Evidence you can inspect

• The working tool can be opened without an account and includes a synthetic agreement for repeatable inspection.
• The client code loads the model only after interaction and performs inference in the browser rather than posting contract text to a SupaMakers analysis endpoint.
• No claim about legal accuracy, time saved, or customer outcome is made without a documented evaluation set and reviewer baseline.

What this demonstrates

ContractScanner demonstrates a local-first AI product pattern: narrow the task, make privacy architectural, expose limitations, and keep the final commercial or legal decision with the user. A production deployment would add a domain-specific evaluation set, approved playbooks, reviewer feedback, and controlled release criteria.