On-premise AI Hub — demo for a major Korean card issuer

Context

The on-premise AI Hub is a citation-backed AI demo that runs with 0 external API calls inside a bank-grade network-segregated environment, answering 11/11 golden-set questions over 877 chunks from 13 actual terms and law documents. We built a demo addressing 5 projects in a major Korean card issuer's AI Transformation roadmap. The financial institution's network-segregated environment required 0 external API calls. Every answer had to cite supporting provisions, and even refused questions had to be included in the audit records.

Problem

With cloud LLMs blocked, AI use had to start from 0. A demo built on synthetic data could not support performance claims, so it had to use actual terms and laws.

What we did

Mapped 5 modules (Doc · KB · Chat · Call · Audit) to the 5 roadmap projects on a 1:1 basis.
Built the corpus from 13 actual public documents containing terms and laws.
Ran answer generation, classification, and speech recognition entirely on local models, and enforced external call blocking with OFFLINE_MODE.
Recorded blocked and refused questions in the audit log with their reasons.
Scored call analysis on compliance with mandatory disclosures and the use of prohibited expressions.

Results

Metric	Value	Notes
Corpus	13 actual terms and law documents · 877 chunks
Accuracy	11/11 golden set questions answered correctly	All citations matched the expected provisions
Response time	6~10 seconds	Local 35B model
External API calls	0 calls
Call analysis	Normal calls scored 92·95, deficient call scored 15	2 missed mandatory disclosures + 2 prohibited expressions detected
Audit records	11/11	Including 4 blocked cases with reason badges

Timeline · Scale · Stack

Timeline: 2026-08 ~ 2026-09 (demo 2026-09-17)
Scale: 5 modules, a single workstation (local inference)
Stack: Local LLM (35B · 3B) · local STT · vector search · launchd persistent services

What this case shows

AI can run in a network-segregated environment. With cloud models unavailable, evidence citations and audit records become central to the design.