'Growth Note' 8-Month Development Story

'Growth Note', a settlement analysis service for Coupang RocketGrowth sellers, automatically analyzes and displays product profitability and key monthly checks when complex settlement Excel files are uploaded.

It was designed to solve the problem of complex settlement data analysis that arises when using Coupang RocketGrowth. An MVP was developed in February based on the settlement master sheet, and it was officially launched in April.

The main goal was to alleviate the inconvenience of sellers who previously had to manually merge up to 40 Excel files of eight different types every month.

Coupang handles the entire logistics process including packaging, delivery, returns, and customer service, but settlement is up to the seller. The analysis gap widens depending on Excel proficiency, and sellers who have difficulty using Excel struggle to distinguish between profitable and loss-making products. Although Coupang handles the entire logistics process from packaging to returns, sales analysis results vary greatly depending on Excel proficiency. Sellers with low proficiency give up on analysis and find it difficult to distinguish between profitable and loss-making products.

This situation causes a significant disparity in settlement analysis results depending on the seller's Excel proficiency. Furthermore, sellers who have difficulty using Excel sometimes give up on analysis entirely, failing to identify which products are profitable and which are incurring losses.

Solving the difficulties sellers faced in manually processing numerous settlement data was the core motivation behind the development of Growth Note.

Even for Excel proficient users using formula templates, it took 1-2 hours per month, while non-proficient users spent over 10-20 hours per month or often gave up on analysis entirely.

However, introducing Growth Note allows settlement and analysis tasks to be completed in just 5 minutes, drastically improving work efficiency.

Growth Note addresses this by introducing 'Red Box' based on tax invoices and 'Blue Box' reflecting actual costs. Sellers enter their costs into the app to clearly understand the net profit per product.

It was difficult to calculate actual operating profit because the data provided by Coupang did not include cost of goods and expenses. However, the Growth Note app allows sellers to input cost of goods and fixed costs for each product and calculate their actual net profit.

This is considered a killer feature that helps sellers bridge the gap between sales revenue on tax invoices and actual operating profit, enabling them to make more accurate business decisions.

Growth Note's original UI/UX was created with a combination of Opus and Astra, and a blue screen was preferred among the warm and clean design options.

However, current satisfaction with the screen layout and font is low, so there are plans to improve the UI/UX by gathering user feedback.

The developer stated that a user vote will be conducted to explore improvement directions.

Growth Note operates by users downloading a manual and uploading Excel files. Analyzing up to 8 types and 40 Excel files was a key technical challenge.

To address concerns about the leakage of personal and business data, a 'temporary analysis' feature was implemented, allowing analysis without storing data in the database.

This feature enables users to utilize the service with peace of mind, without security concerns.

Verifying data consistency is the primary goal and raison d'être of this service. The dashboard displays Coupang data in a 'Red Box' and calculated operating profit in a 'Blue Box', helping users grasp their actual net profit. This allows for revising sales strategies by understanding the profit structure of each product.

Verifying data consistency is the primary goal and raison d'être of the service. The dashboard allows immediate confirmation of sales and profit per product, clearly distinguishing between profitable and loss-making items.

The analysis model utilizes the latest high-performance AI, such as Opus 5.5. Based on monthly sales data, it checks for missing purchase prices and suggests ways to allocate advertising costs or optimize margins. Specific action plans, such as discontinuing unprofitable options or strengthening advertising for high-margin items, are also provided.

It analyzes monthly sales data to propose advertising budget allocation and margin optimization, providing guidance such as consolidating ad spend to a monthly budget of 8 million won and reallocating it mainly to high-margin options.

It also checks for and notifies about missing input data, such as unentered purchase prices, and provides specific execution plans like streamlining unprofitable options and strengthening ads for high-margin options, helping sellers apply them to their actual businesses.

The implementation of MCP (Model Context Protocol), which serves as a link between agents and services, is urgent. The service must be linked to various environments such as Claude or Codex, and page design that allows users to connect directly or use prompts is essential.

MCP (Model Context Protocol) acts as a link between agents and services, and the service must be designed to interoperate with various environments such as Claude or Codex.

A page where users can connect directly or utilize prompts is essential, and this is considered a critical factor for the service to maintain competitiveness in the future AI ecosystem.

Upload cards within the service act as an intermediary between the agent and the service, helping users easily use the service through the agent.

The developer explained that this significantly reduced the psychological barrier (friction) users experience when using the service.

Coupang seller optimization know-how secured through expert interviews was reflected in the logic. Beyond simple feature implementation, accumulated domain knowledge has become a competitive advantage that differentiates it from general AI analysis.

During the development process, in-depth interviews with experts helped incorporate Coupang seller optimization know-how into the service logic, which forms the basis for providing business insights necessary for Coupang sellers' business optimization.

This accumulation of domain knowledge goes beyond simple feature implementation and acts as a crucial factor in building Growth Note's core competitiveness, or 'moat'.

Growth Note's service architecture utilized Vercel deployment for the frontend and Fast API for the backend to maximize efficiency.

Particular emphasis was placed on parsing tasks, which involve many calculations and require memory optimization. The core goal was to implement the process from Excel upload to analysis and AI insight provision as quickly as possible.

Supabase was used for database and authentication, ensuring both development speed and efficiency.

Growth Note introduced a one-time temporary upload feature to alleviate user reluctance to upload sensitive business data.

This feature lowers concerns about personal data leakage by analyzing user data without storing it in the database.

While accumulating history is necessary to improve the accuracy of AI insights, information stored in the database is masked so developers cannot access its content. Security-sensitive users can use only the temporary upload feature for analysis without storing history.

Securing a browser end-to-end (E2E) testing structure was essential for rapid application verification and debugging. In the early development stages (January-February), when AI browser testing performance was insufficient, the focus was on implementing using Playwright MCP.

At the time when Claude AI first emerged, browser testing was very inadequate, yet it was essential to establish a browser end-to-end testing structure for rapid application verification and debugging.

The developer stated that E2E testing significantly contributed to securing system stability and verifying the accuracy of complex logic.

Communication delays due to time differences and a lack of understanding of the business ecosystem acted as a more significant bottleneck in the development process than coding. Much time had to be spent on data verification without domain knowledge. Growth Note found solutions by sharing screenshots or entire drive folders using MCP when AI couldn't understand the specifications. In summary, the bottleneck was not the code but the people, and the process of building domain knowledge and context to determine data consistency was the problem, according to an official.

Communication delays due to time differences and a lack of understanding of the business ecosystem were major challenges in the development process. In particular, much time was spent on verification without domain knowledge to determine if the numbers were correct.

When AI failed to properly read the project documents, solutions were sought by taking screenshots or sharing entire drive folders using MCP.

In the early stages of AI service development in January 2023, there was an experience where AI, at a junior engineer level, had plausible designs but many bugs in reality.

In particular, a lot of time was spent on testing and verification when the developer lacked domain knowledge, which was a major challenge in securing service stability.

As a result of clearly setting end-to-end testing metrics and repeatedly verifying based on them, service stability was secured, and the number of bugs was significantly reduced.

Creating an environment where agents immediately respond to data and accept feedback is crucial for service operation. It is considered much more effective to build custom telemetry optimized for service requirements than using existing tools like Google Analytics. Even if maintenance costs arise, custom building is recommended for service freedom and insights, it added. It emphasized that custom building can be a better choice in a highly flexible, modern agent environment.

With the advancement of AI agents, custom telemetry enables the creation of an environment where agents can immediately react to data and incorporate feedback.

Even if maintenance costs are incurred, custom construction is strongly recommended to maximize service flexibility and operational insights.

The developer emphasized that if they were to build the service again from scratch, they would design it exclusively for AI agents.

This demonstrates the importance of an agent ecosystem where AI connects with external services, much like an app store, and highlights the need for agents within the service to bridge the gap between humans and AI.

The judgment is that only by designing a structure where agents can efficiently use the service, such as MCP (Model Context Protocol), can the service be competitive in the future AI environment.