What to do in such situations: Developing an app integrated with generative AI
Hello, I’m pj-corridor, an engineer. In this article, I will introduce the challenges I faced while developing an application integrated with generative AI, such as dealing with abuse and navigating trial-and-error and trade-offs in prompt tuning. I previously had the opportunity to work with Synergy Marketing, and I am grateful that they kindly agreed to publish this article on the TECHSCORE…
In this article, pj-corridor, an engineer, shares their experience developing an application integrated with generative AI, such as personality assessment tools. They highlight the challenges faced during the development process, including dealing with abuse and navigating trial-and-error and trade-offs in prompt tuning. The author previously worked with Synergy Marketing and is grateful for their consent to publish this article on the TECHSCORE BLOG.
The app provides CAPS and DiSC personality assessments, which are pseudoscientific tools meant for self-introduction sessions rather than decision-making. The author uses the app for such purposes, such as workshops, to energize participants.
However, generative AI has its quality risks, as it can introduce hallucinations. To mitigate these risks while keeping the cost of ensuring quality within the scope of a hobby project, the author kept the rule-based scoring logic deterministic and limited the generative AI output to supplementary text.
The article is divided into three chapters: abuse prevention, technology selection, and trial-and-error examples. In the abuse prevention chapter, the author discusses strategies to handle bot abuse, such as using a Web Application Firewall (WAF) and controlling access from simple bots and browsers. They opted to use a Lambda function URL for deployment, which requires AWS-specific implementations like changing the authentication type and managing environment variables.
However, they decided against implementing these measures temporarily due to the risk of malicious actors discovering the Lambda Function URL and bypassing the WAF.
The technology selection chapter covers the author's trials and errors with Lambda and API Gateway. Initially, they preferred using a Lambda container image, but eventually chose to use a Lambda zip deployment instead, as it provided better compatibility with AWS and reduced complexity in their continuous integration pipeline. They also emphasize the importance of promptly cleaning up unused resources to avoid technical debt.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
