An AI engine for classifying open-ended survey answers

A platform automating the most tedious job in quantitative research — coding thousands of open-ended answers — with a 4-phase LLM-based algorithm, used daily in production by research teams.

Challenge

In quantitative research, open-ended answers (“Why do you say that?”) must be classified into categories — days of manual work for thousands of responses, prone to coder inconsistency, and a bottleneck in nearly every project.

Approach

I built an automatic classification platform with a four-phase algorithm: candidate category extraction, LLM-based semantic merging (the breakthrough phase — similar categories are detected by meaning, not word similarity), full classification, and quality control.

The design is grounded in research reality:

Supports semi-open questions (pre-coded categories + “Other”) which dominate real questionnaires.
Multi-tenant with OTP registration — one installation serves many teams.
Costs are tightly managed: economical model selection brings costs to roughly $0.15 per million tokens.

Result

Live in production and used daily by research teams: an average of ±60 seconds per variable at 100% response coverage — work that used to take days now finishes in minutes, with auditable category consistency.