Skip to main content
Back to the villageDONGGU’S FIELD NOTESGS-CALTEX-VOC-AI-DASHBOARD
The workshop · selected work
PLACE
GS칼텍스
ROLE
Full-Stack
STARTED
2024-10
★ Case 01 / 5Forward Deployed Engineer Donghyun Kang

GS Caltex VOC AI Dashboard

Took an environment where VOC piled up as dead data and built it into a unified analysis system from zero.

GS칼텍스Full-Stack2024.10 ~ In progress
01.Discover

It was dead data that never arrived anywhere.

The mission was one line. Once I looked inside, classifier accuracy was actually the smallest problem — four channels (app, call, survey, social) were piling up in different places, junk like "hhhhhh" was flowing through to the classifier, and the CS team was reclassifying thousands of items every week by hand, paying for the missing system with people.

The job was not to build one good classifier. It was to stand up the full loop — ingest, clean, classify, act — from zero. That was where the project actually started.

Field notes

Note 01

There was no unified ingest

The four channels — app, call, survey, social — were piling up in separate places.

Note 02

There was no gate in front of the classifier

"hhhhhh", "lol"-equivalents, and empty strings flowed straight through into the 60 categories.

Note 03

People were holding the system up

CS was reclassifying thousands of items every week by hand.

02.Build with AI

Used AI as a thinking tool, not as the model.

Decision 01

Compare and swap the model

BeforeGPT-4: expensive; 70% accuracy on 60 categories
AfterClaude Haiku + prompt tuning: lower cost; 88% accuracy

tradeoffKorean-topic accuracy vs. API cost

Decision 02

Prompt engineering

A "neutral classifier persona" system prompt blocked LLM hallucination. A gaslighting-style pattern broke any carryover from the previous turn.

Decision 03

Preprocessing rules

Junk text was removed before it reached the LLM, using regex and length rules. Cut unnecessary API calls by 30% and recovered more accuracy on top.

Decision 04

RAG pipeline

Vector embeddings plus topic matching restructured the 60 categories into 6 intents × 10 sub-categories. Turned it into a searchable analytics asset.

VOC AI dashboard, automatic topic classification
03.Prove with metrics

Only the results you can read off an instrument.

Manual → automated

0 → 1000+

CS classifications / day

Topic accuracy

70 → 88%

+18pp

Auto-ingest channels

4 channels

App · call · survey · social

API cost

−30%

Junk blocked by preprocessing rules

What this case is really about
“An engineer who picks the metrics, owns the cost-vs-accuracy tradeoffs, and uses AI as a thinking tool.”
Discover

Read 60-category ambiguity as a pattern out of the live ops data.

Build

Compared and swapped LLMs, tuned the prompts, lifted accuracy by 18 points.

Prove

CS hand-tagging went from zero to 1000+ classifications per day; API bill dropped 30%.

Linked principlesEvidence-DrivenOwnership