MeiryoAI

What inspired you to build MeiryoAI?

We noticed a significant shift in how the 18-35 demographic searches for information. Increasingly, they're turning to Gemini, ChatGPT, and other AI platforms in addition to traditional search engines. This is the exact demographic that universities need to reach for undergraduate and graduate recruitment.

This reminded us of the early 2010s when universities realized they needed to manage their presence on social media platforms. We view AI search as the next important channel, but universities have no tailor-made way to monitor how they appear in these AI-generated answers.

What made this challenge compelling is that universities face unique reputation challenges. We wanted to build an analytics tool specifically tailored to higher education that solves these distinct challenges, rather than an AI analytic solution that tries to serve every vertical.

What did you learn throughout the building process?

We learned that building effective analytics requires constant iteration.

Lesson 1 - Visibility Scoring

Initially, we thought brand visibility was a binary score: either a school appears (100%) or it doesn't (0%). This gave school marketers no actionable insights. In version 2, we introduced a weighted scoring system that considers:

Base score – presence/absence
Frequency score – how often mentioned
Depth score – how substantively discussed

This nuanced approach gives users meaningful data they can actually act on.

Lesson 2 - Priority Scoring

Our first position score formula was mathematically correct but not actionable. It clustered all scores between 9 and 10. For example, a school mentioned at character 500 in a 10,000-character response scored 9.95, which told users nothing meaningful.

Version 2 addressed three improvements:

Full scoring range – Using weighted position ratios that leverage the entire 0-10 scale
Multiple mentions – Accounting for schools mentioned 10+ times, not just first appearance
Distribution value – Recognizing that early + distributed mentions are more valuable than a single late mention

The new formula:

Priority Score = Placement Score × (1 + Distribution Bonus) Placement Score = 9.0 × (1 - weighted_position_ratio) Weighted position ratio = (first_position × 0.7 + average_position × 0.3) / total_words

Distribution Bonus divides responses into 10 sections and rewards mentions across multiple sections

Lesson 3 - Google Gemini API

We also obtained new skills about:

JSON-formatted responses for structured data extraction
Grounding capabilities to pull real-time web information
Citation and source extraction from AI answers
Using Gemini as a transformer for deep sentiment analysis and generating actionable insights

How did you build MeiryoAI?

Design first – Sketched dashboard elements in notebooks, visualizing what insights users would need
Logic before code – Documented analytical logic on paper before writing any code
Backend validation – Ran backend code for several days to ensure stability and accurate data collection
Frontend development – Built data visualizations and dashboard after backend validation
Iterative refinement – Fixed UI bugs one at a time, anticipating end users’ needs and clarity of presentation

What challenges did you face?

Frontend Development

First time using Vite + React together (previously only React)
Not being a frontend developer made creating a compelling, user-friendly UI particularly challenging
Improved knowledge of modern tooling and design patterns while building

Backend Development

Defining meaningful metrics – Initially struggled with what to analyze from Gemini answers. Iterated to find metrics that help university marketers
Memory optimization for deployment – Full dataset overwhelmed server memory on Render. Optimized by processing 2-3 user accounts concurrently instead of all at once
Security improvements – Migrated from tokens in local session storage to cookies for better protection against attacks

Architecture Design

Prompt redundancy problem – Initially nested prompts under each school, which required duplicating prompts for multi-school analysis. Restructured data architecture so prompts and schools work independently but can be combined during analysis