Building a Sales Prep AI That Went Deeper Than Expected
The author built a sales prep AI tool that automatically researches a company and generates a report before a sales meeting. The process involved several challenges, including handling long processing times, migrating from Slack to LINE, and addressing AI hallucinations.
Why it matters
This project demonstrates the challenges of building a practical AI-powered tool and the importance of addressing issues like performance, integration, and model reliability.
Key Points
- 1Developed a sales prep AI tool that automates company research and report generation
- 2Faced issues with Vercel's 10-second timeout and solved it using streaming responses
- 3Migrated the tool from a Slack bot to a LINE bot to improve accessibility
- 4Struggled with the AI generating plausible but inaccurate information (hallucinations)
Details
The author started with a simple web form that allowed users to enter a company name, department, and contact name, and the tool would search the web, analyze the results using Claude AI models, and return a report. To address the Vercel timeout issue, the author implemented a streaming response approach that kept the function alive during the 1-2 minute research process. The tool was initially built as a Slack bot, but when the author's paid Slack plan was canceled, they migrated it to a LINE bot. The biggest challenge was fighting AI hallucinations, where the model would confidently return plausible but inaccurate information. The author had to work on improving the reasoning quality and fact-checking capabilities of the AI.
No comments yet
Be the first to comment