ScopeBack: An AI Tool to Find Hidden Scope Creep in Freelance Projects

Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle Submission 💸

A freelance project starts with a quote, a list of deliverables, and a price everyone agrees on.

"Can we add one small section?" "Could you also change how this form works?" "One last revision, please."

Individually, these requests don't seem like much. But across a project, they can turn into hours of extra work that are difficult to track, document, or discuss with a client.

I started building ScopeBack, a local-first AI tool that helps freelancers compare their original project scope with client conversations and identify potential scope changes worth reviewing.

The goal is simple: help freelancers understand where extra work may have happened, with evidence rather than guesswork.

ScopeBack is an MVP designed for freelance developers, designers, and small agencies that manage client projects through WhatsApp and email.

The last step is intentional. ScopeBack doesn't assume that every additional request was unpaid, completed, or billable. A request might already have been approved, included in the agreement, or settled with the client.

The AI highlights what deserves attention; the freelancer makes the final decision.

The current MVP supports text-based PDF, TXT, and Markdown inputs where applicable, includes a synthetic sample project, and can export the reviewed findings as a Markdown report.

The current version is designed to run locally.

The recording will show the complete workflow: importing a sample quote and conversation, analyzing the requests, reviewing the evidence, and generating an audit report.

The project includes a clearly labelled sample report that can be explored without running an AI model.

GitHub: https://github.com/DiyaVj/scopeback

The repository includes the FastAPI backend, browser interface, document and conversation parsing, analysis pipeline, evidence validation, sample project, and automated tests.

python -m venv.venv source.venv/bin/activate pip install -r requirements.txt cp.env.example.env ollama pull qwen3.5:9b ollama serve Enter fullscreen mode Exit fullscreen mode In another terminal, start the app:

uvicorn app.main:app --reload Enter fullscreen mode Exit fullscreen mode Open http://127.0.0.1:8000 in your browser.

The README contains the full setup instructions and configuration options.

The AI component is designed around local inference with an open-weight language model. The default configuration uses qwen3.5:9b, which can be replaced with another compatible model.

I wanted the model to return structured findings instead of a long, unstructured response.

The analysis pipeline extracts the project's agreed scope, processes the client conversation, assigns canonical IDs to individual messages, and uses those IDs when referencing potential scope changes.

The backend then validates the returned evidence references against the actual conversation. This helps prevent fabricated message references from appearing in the report.

I also kept the financial calculations separate from the AI. The model doesn't decide how many hours a request took or invent a price. The user supplies those inputs and confirms which items should count toward an estimate.

The MVP includes automated tests for parsing, API behaviour, and evidence grounding. Real-world accuracy and the rate of false-positive findings still need further evaluation.

For this project, open-weight AI isn't just a way to add an AI feature. It shapes how the product can work.

Quotes and private client conversations can contain commercial information. With the default local configuration, ScopeBack sends analysis requests to an Ollama instance running on the user's machine rather than to a hosted LLM API.

That gives freelancers more control over their data. The app doesn't need a database to save uploaded documents in this MVP.

Once the model is downloaded, inference can run locally without a separate per-request fee from a hosted model provider. There are still hardware, electricity, and setup costs, and performance depends on the machine being used.

For a small side project aimed at cost-conscious freelancers, that trade-off is worth exploring.

The analysis pipeline can be tested with different compatible models. This makes it possible to compare evidence accuracy, latency, and resource usage instead of assuming that one model is always the best fit.

4. The product can be built around a specific workflow.

Rather than asking a general-purpose chatbot to summarize a conversation, ScopeBack structures the task around an agreement, individual requests, source-message evidence, human review, and user-confirmed estimates.

The interesting question isn't simply whether the model can understand a conversation. It's whether it can make the review process useful without creating more work for the freelancer.

The next milestone isn't adding more features. It's validating the problem.

I want to test ScopeBack with freelance developers, designers, and small agencies using anonymized project examples, with permission from the people involved.

I also want to measure false positives, because an inaccurate report could create unnecessary friction between a freelancer and their client.

ScopeBack is an early MVP, not a legal or accounting tool, and it doesn't establish that any particular work is recoverable revenue.

The longer-term goal is to find out whether a small, privacy-conscious AI tool can solve a sufficiently painful problem to become a sustainable side hustle.