Inspiration
In China, many freelancers have relatively little bargaining power compared with their clients. Competition is intense in many industries, and freelancers often depend heavily on the income from each project.
Video editors working for content creators, design companies, and other clients frequently face overwhelming workloads. This is often not because they are unable to complete the creative work, but because one seemingly ordinary and informal message quietly turns into yet another revision round. Meanwhile, the originally agreed project scope, excluded work, and pricing information may be scattered across chat histories and different documents.
Before moving into the AI industry, I worked as a geological engineer at a copper mine in Afghanistan. I shared information I had learned about Afghanistan's gold mining industry on social media. This attracted the attention of several commercial research organizations that urgently wanted to understand Afghanistan's mining industry and mining supply chain, and they commissioned me to conduct research.
Because a research report can only use a limited amount of text and images to represent the conditions of an entire industry, this side job exposed me to the same problems of changing requirements, mutual distrust, and disagreements over whether the information provided was valuable. I made more than twenty revisions according to their requests, but they still tried to reduce the agreed price by claiming that the information did not meet their requirements.
I also have a friend who works in the design industry and frequently provides services such as video editing, photo retouching, and logo and trademark design. She often complains to me about conflicts between clients and service providers over design changes and the additional fees associated with those changes.
These experiences inspired me to build CutClause. Its purpose is to make the commercial boundary clear before additional work begins.
The same workflow can also help freelancers in many other fields, not only video editors. Anyone facing a similar pattern of repeated changes can use the logic of the app to define the price of additional revisions and obtain explicit confirmation before continuing the work.
What it does
CutClause is an approval-gated AI operations agent for freelance creative projects.
A freelancer begins by entering an informal client brief. The Scope Agent converts it into a structured project proposal that includes:
- Included deliverables
- Excluded work
- Included revision rounds
- Project price
- Price for additional revisions
- Delivery assumptions and potential risks
The freelancer must review the proposal before applying it to the project. The client can then use an account-free link to submit one consolidated batch of time-coded revision requests.
Whenever a revision request is received, CutClause organizes and normalizes the requested changes, compares them with the approved project agreement, and determines whether the work remains within the included scope.
When all included revision rounds have been used, the next revision becomes a clearly visible quote and approval checkpoint instead of silently turning into additional unpaid work.
CutClause prepares the recommended next action, but irreversible commercial decisions remain under human control. The agent cannot independently modify the project agreement, send messages to the client, initiate a charge, or claim that payment has been received without supporting evidence.
What makes it agentic
CutClause is a stateful workflow rather than a single chat response.
A client feedback event moves through the following observable and auditable execution events:
revision.submitted
→ feedback.normalized
→ scope.checked
→ extra_revision.quoted
→ freelancer_approval.requested
These events mean:
- The client submits a revision request.
- The agent converts the feedback into structured information.
- The agent compares the feedback with the approved project scope.
- If the request is outside the included scope, the agent generates an additional-revision quote.
- The agent requests explicit human approval from the freelancer.
The agent considers the approved project terms, the number of revision rounds already completed, the structured feedback, and the workflow policy to determine which action is permitted next.
Human approval checkpoints protect changes involving project scope, pricing, client communication, and payment-related states. This prevents the agent from independently making irreversible commercial decisions.
The interface also clearly distinguishes between real production execution and local demonstration results:
live_agent: The request was actually processed by the deployed production agent.transparent_demo: The interface is displaying a local preview or fallback demonstration result, and no real model call occurred.
This distinction prevents demonstration data from being mistaken for a real production model result.
How I built it
The production agent service was built with Google Agent Development Kit and uses Gemini 3.5 Flash through Vertex AI.
The agent service is deployed on Google Cloud Run. Firestore stores agent execution records and decision evidence.
The client application was built with Expo, React Native Web, and TypeScript. Supabase Auth and PostgreSQL support creator accounts, project records, public client links, and structured client feedback history.
The production workflow includes the following steps:
- Extract the project scope from an informal creative brief.
- Perform structured validation of deliverables, exclusions, pricing, and risks.
- Require human review before the proposed project terms are applied.
- Collect time-coded feedback through a link that does not require the client to create an account.
- Check the remaining revision allowance and identify out-of-scope work.
- Require explicit human approval for additional work.
- Handle the payment handoff as a separate workflow event without displaying an unverified payment as completed.
Challenges I faced
The biggest challenge was determining how freelancers and AI agents could work together effectively.
I left my job as a geological engineer and entered the AI industry because I want to build AI-era infrastructure for the rapidly growing population of freelancers and startup founders. My goal is to help individuals create business models that are more flexible, convenient, efficient, and economically sustainable.
A useful operations agent needs enough autonomy to help this growing group of people earn real income. At the same time, it should also help clients who are willing to pay for high-quality work access a broader range of services that better match their needs.
The challenge is to create an agent that can actively organize information and prepare useful actions while keeping important commercial decisions, such as scope, pricing, approval, and payment, under human control.
What I learned
Running a real business is difficult and inefficient, and it requires a great deal of time.
While building this app, I took two or three wrong turns and spent much more time than I initially expected. If building and operating a product remains this difficult even with the assistance of AI, operating a company in the real world is even more challenging.
Running an app requires substantial effort and cost, and the barrier to entry can be high. This experience strengthened my desire to build more infrastructure for the AI era so that startup founders and freelancers can work more efficiently and move closer to economic freedom.
I also learned that:
- An account-free client link reduces friction for clients who need to participate in the workflow.
- Revision requests are easier to manage when they are submitted as one structured batch instead of being scattered across chat messages.
- Approval and payment must be treated as two separate workflow events.
What's next
Next, I plan to add:
- Consent-based notification features
- Payment-provider-verified webhooks
- More complete agent execution history
- Reusable project agreement templates
- Real pilot testing with freelance creative teams
- Richer analysis of project scope and revision costs
Demo-data disclosure
NOVA Launch Film, Maya Chen, Alex, all client feedback, prices, approval states, and payment states shown in the demo are synthetic test data.
This project does not claim that any real customer transaction occurred, and it does not claim that any payment shown in the demo was completed.
Built With
- adk
- ai
- cloud
- expo.io
- fastapi
- firestore
- gemini
- native
- postgresql
- python
- react
- run
- supabase
- typescript
- vertex
- web
Log in or sign up for Devpost to join the conversation.