ScopeGuard: Helping Freelancers Handle Scope Creep with Open-Weight AI

This account chronicles the development of ScopeGuard, an innovative web application designed to assist freelancers in managing scope creep in client relationships. The tool addresses a common challenge faced by freelancers: clients often request additional features, revisions, or deliverables beyond the original agreement. The issue isn't solely in deciding whether to accommodate the request, but rather in finding a way to respond professionally while preserving the client relationship.

ScopeGuard aims to alleviate this dilemma by analyzing new client requests against the agreed-upon project scope and guiding freelancers on how to respond.

ScopeGuard's functionality is straightforward yet powerful:

1. Defining the Scope: Users can meticulously record the project's deliverables, features, exclusions, and relevant agreement terms.

2. Analyzing Client Requests: The application breaks down new requests into individual items and meticulously compares them against the provided scope.

3. Reviewing Evidence: Supporting quotes are inspected to identify potential additions or ambiguous requirements.

4. Drafting a Reply: ScopeGuard generates a professional response, negotiating additional work, seeking clarification, or declining the request as needed.

What sets ScopeGuard apart from generic AI email writers is its insistence on grounding its analysis in the original agreement. This focus on agreement validation sets it apart from other tools. The application is particularly relevant for freelance developers, designers, and small agencies that frequently deal with changing requirements. A potential business model includes a free tier with limited analyses and a paid tier for more extensive usage.

The development of ScopeGuard was built using a combination of technology:

- React, Vite, and TypeScript for the frontend.

- Chakra UI for interface components.

- FastAPI and Python for the backend.

- Ollama for local model inference.

- Qwen2.5 3B as the default language model.

A significant challenge in using language models is their propensity to make claims unsupported by the supplied agreement. To mitigate this, ScopeGuard integrates application-level validation:

- The request is evaluated in smaller parts.

- Structured output is validated against the expected schema.

- Supporting quotes are checked against the original scope text.

- A missing or invalid quote prevents the application from treating an unsupported claim as verified evidence.

The final classification is computed in application code, explicitly rejecting unsupported conclusions from the model. Although this approach cannot eliminate all model errors, it significantly reduces them. The decision to use a small open-weight model was driven by the desire to explore whether a useful workflow could run locally without relying on a hosted LLM API for every analysis. While smaller models may struggle with complex agreements and ambiguous language, users must still review the results.

Open innovation played a crucial role in ScopeGuard's development. By utilizing an open-weight model through Ollama, the application enables local inference, keeping all data processing on the user's machine. This approach not only respects the confidentiality of sensitive business information but also makes experimentation more accessible.

Users can compare different models, inspect their behavior, modify the validation logic, and adapt the application to their specific needs. Open innovation allowed for the exploration of locally run models and the implementation of application-level safeguards, demonstrating the power of open-source solutions in addressing real-world challenges.