LIVING IN THE DATABASE
← Back to Talks
The Problem
Market researchers hate dashboards. They need answers, not visualizations they have to decode. Traditional anime market research tools require SQL knowledge, dashboard navigation, and hours of manual cross-referencing between demographic data and anime metadata.
FanServ US, a market research firm specializing in fandom demographics, had thousands of survey responses about anime preferences — but no way for non-technical researchers to query that data alongside MyAnimeList's comprehensive anime database. The barrier to insight wasn't the data; it was accessibility.
What We Built
A multi-agent AI research tool that lets market researchers ask questions in plain English and get actionable insights from 60,000+ anime titles combined with FanServ's proprietary demographic survey data.
The Three-LLM Architecture
Instead of a single AI attempting to handle everything, we built three specialized agents, each with a distinct responsibility:
- Translator Agent: Converts natural language queries into SQL. Understands industry terminology and anime-specific concepts. Handles complex queries involving demographic filters, genre combinations, and temporal analysis.
- Auditor Agent: Validates SQL queries before execution. Checks for logical errors, performance issues, and potential data quality problems. Prevents expensive queries and ensures accurate results.
- Investigator Agent: Identifies data gaps and suggests follow-up queries. Detects when demographic segments are underrepresented or when temporal patterns need deeper analysis. Proactively surfaces research opportunities.
The Data Layer
Dual database architecture combining PostgreSQL for structured demographic and anime metadata with a graph database for relationship mapping (studios, creators, franchises, genre evolution). 60,000+ anime titles from MyAnimeList integrated with FanServ's proprietary survey data covering demographics, viewing habits, and brand affinity.
Live Demo Examples
During the panel, we demonstrated three real-world market research queries that showcased the system's capabilities:
Query 1: Chiikawa Brand Metrics
Question: "What demographic segments show the highest affinity for Chiikawa-style slice-of-life content, and what are the brand partnership opportunities?"
The system identified Gen Z and millennial women as primary audiences, cross-referenced with brand affinity data to surface stationery, lifestyle, and fashion partnership opportunities. The Investigator agent flagged that international viewer data was limited and suggested follow-up demographic expansion.
Query 2: Yoko Kanno Analysis
Question: "Which anime soundtracks by Yoko Kanno correlate with the highest rewatchability metrics, and what does that tell us about music-driven viewer retention?"
The system pulled Cowboy Bebop, Ghost in the Shell: Stand Alone Complex, and Terror in Resonance, correlating high rewatch rates with jazz and electronic scores. The Investigator suggested examining streaming playlist behavior as a follow-up retention signal.
Query 3: Emerging Genre Detection
Question: "What new genre combinations are gaining traction among under-25 audiences in the last 18 months?"
The system detected rising interest in psychological isekai, romance-thriller hybrids, and workplace comedy-dramas. The Auditor flagged that sample sizes for workplace comedy-drama were statistically small, recommending cautious interpretation.
The Room
Standing room only. We expected a modest crowd for a data-focused panel at an anime convention, but the room filled 30 minutes early. Industry researchers from Crunchyroll, Aniplex, and independent market research firms were in attendance. The Q&A went 20 minutes over time.
The most common question: "When can we use this?" followed by "How do we integrate our own data?" The reception validated what we suspected — the anime industry is starving for accessible research tools that don't require a data engineering team.
What's Next
This was a CMU MSBA capstone project, but the work continues post-graduation. FanServ is exploring commercial licensing for their client base. We're working on open-sourcing the multi-agent framework (without proprietary data) for other research domains.
Gem Partners, a Japanese anime market research firm, has expressed interest in a localized version for the Japanese domestic market. Conversations are ongoing.
If you're a market researcher, data engineer, or fandom analyst interested in collaborating or deploying this system, reach out.