ScopeGuard: How I Built an NDA-Safe, Local Gemma 3 Agent That Saved $1,190 in Freelance Scope Creep

The submission targets freelancers and small agencies that experience uncontracted work requests, often resulting in significant time and money loss. The $19/month micro-SaaS, ScopeGuard, serves as a contract shield for these professionals. When a client sends an extra request, the user drops the message into ScopeGuard, which retrieves the signed Statement of Work (SOW) from MongoDB Atlas and evaluates it using a fine-tuned Google Gemma 3 4B model.

The tool calculates the potential unbilled revenue at risk and drafts a diplomatic change order email complete with itemized line items and a one-click approval quote.

Closed frontier AI APIs like OpenAI GPT-4o and Anthropic Claude 3.5 fail to meet three business requirements: strict client NDA compliance, niche fine-tuning, and cost-effective unit economics. ScopeGuard, on the other hand, runs open-weight Google Gemma 3 4B locally, ensuring that zero bytes of client IP ever leave the infrastructure.

The tool also employs a specialized fine-tuned Tinker LoRA adapter trained on freelance SOW clauses and conversational scope-creep scenarios, which outperforms base models at detecting contractual boundary violations.

When a client requests work outside the agreed scope, ScopeGuard flags it as OUT_OF_SCOPE_CREEP, citing the relevant SOW section, calculating the revenue at risk, and generating a polite change order email. The tool provides a live dashboard for analyzing incoming client messages and SOW clauses, complete with risk scores, protected revenue calculations, and generated emails.

The entire workflow is powered by a stack of partner technologies, including Mastra AI Agent Engine, MongoDB Atlas, Sentry Distributed AI Agent Tracing, Render, and Tinker LoRA.