Self-Hosted AI Deployment in Seoul, South Korea

Financial Services / Credit Card / Self-Hosted AI / Seoul

A Seoul-based Korean financial group's card and payments division (KRW 4.2T revenue — approximately USD 3.2B, 18 million active cardholders, processing 2.8 billion annual transactions, 4,800 employees) needed to deploy AI capabilities to transform fraud detection, merchant analytics, and customer service. The division had invested in traditional ML models for years (gradient boosting fraud models, logistic regression credit scoring), but the next generation of capabilities — LLM-powered customer service, complex fraud pattern recognition, and merchant intelligence — required self-hosted LLM infrastructure. Core challenges: (1) FSC data residency requirements — the division processed: 18 million cardholders' personal information (names, addresses, phone numbers, birth dates — PIPA-regulated), transaction data (2.8 billion transactions annually — credit card transaction data classified as "credit information" under the Credit Information Act, requiring domestic processing), merchant data (850,000 affiliated merchants — business financial information), and card application data (including income information, employment details, and credit bureau data — the most sensitive category under the Credit Information Act). The FSC's cloud computing guidelines classified transaction data and credit information as "critical data" requiring Korean-resident processing. The FSS conducted annual IT examinations specifically reviewing how the division handled critical data — deploying AI on US cloud APIs would create immediate examination findings. (2) Real-time fraud detection — the division processed 7.8 million transactions daily (average, with peaks of 18 million during holiday shopping events like Chuseok and Lunar New Year). Existing fraud models (gradient boosting) processed transactions in 15ms per transaction but had: a 0.42 percent false positive rate (326 legitimate transactions flagged as fraud daily per 1 million transactions — each false positive creating customer friction when cards were declined), a fraud detection rate of 89.2 percent (meaning 10.8 percent of fraud passed through — KRW 82B in annual fraud losses), and no ability to detect complex fraud patterns requiring contextual understanding (social engineering fraud where criminals manipulated cardholders into making legitimate-appearing transactions, account takeover fraud involving subtle behaviour changes, and merchant collusion fraud where affiliated merchants processed fraudulent transactions). LLM-enhanced fraud detection could analyse transaction context, customer communication patterns, and merchant behaviour to catch sophisticated fraud that statistical models missed — but processing 7.8 million daily transactions through LLM inference required on-premise GPU infrastructure, not cloud API calls that would cost approximately KRW 14B annually at that volume. (3) Korean-language customer service — the division handled 1.4 million monthly customer contacts. Korean customers expected: instant response (Korean digital service expectations — influenced by KakaoTalk instant messaging culture — expected sub-3-second response times), natural Korean conversation (not translated-sounding AI responses — Korean consumers quickly detecting and rejecting non-native Korean language), understanding of Korean financial concepts (specific Korean card concepts: "family card," "check card" versus "credit card," "installment payment" — halbu, "point conversion"), and cultural sensitivity (Korean honorifics — jondaenmal — required in customer service, age-appropriate language patterns, and Korean cultural norms around financial discussions). Existing AI chatbot (rule-based) handled 28 percent of inquiries but with 62 percent customer satisfaction — significantly below the 85 percent target. The division needed LLM-powered customer service with native Korean capability. (4) Merchant intelligence — the division had relationships with 850,000 merchants. Merchant analytics currently relied on transaction volume and basic categorisation. The division wanted AI-powered merchant intelligence: merchant risk scoring (identifying merchants likely to commit fraud, go bankrupt, or violate terms — processing financial patterns, transaction anomalies, and public information), merchant recommendation (suggesting new merchant partnerships based on cardholder spending patterns and geographic analysis), and merchant pricing optimisation (dynamic merchant discount rates based on transaction volume, risk profile, and competitive landscape). Merchant data — individual merchant financial information — was commercially sensitive and could not be sent to external providers. (5) Korean-language model quality — Korean is a morphologically complex agglutinative language where meaning changes significantly with particles, honorific levels, and context. Korean financial terminology is particularly complex, mixing native Korean terms with Sino-Korean (Chinese-derived) terms and English loanwords — "credit card" is "sin-yong ka-deu" (transliteration) but many financial terms have both Korean and English forms used interchangeably. Existing open-weight models (Llama, Mistral) had Korean capability but significantly lower quality than English — requiring substantial fine-tuning on Korean financial corpora to reach production quality.

We deployed a self-hosted AI platform over 12 weeks — optimised for Korean financial services with native Korean language capability. (1) Korean-resident GPU infrastructure. Hardware: 32 NVIDIA H100 80GB GPUs across 4 server nodes, deployed in the financial group's primary data centre in Pangyo Techno Valley (the Seoul metropolitan area's technology hub, hosting many Korean financial and technology companies). Configuration: 8x H100 per node, 1.5TB system RAM, 24TB NVMe storage, InfiniBand HDR200 interconnect. The financial group's data centre met FSC/FSS requirements for critical data processing — Tier 3+ reliability, physical security, and access controls verified during annual FSS IT examination. Network: connected to the division's internal network, payment processing network, and fraud detection systems. No external internet connectivity for the AI cluster — all processing within the financial group's secure network boundary. (2) LLM-enhanced fraud detection: catching what statistics miss. Architecture: two-tier fraud detection system combining existing statistical models (fast, processing every transaction) with LLM analysis (deep, processing flagged transactions). Tier 1 — statistical models (existing): gradient boosting models processing every transaction in 15ms — scoring transactions on 280 features (amount, location, time, merchant category, velocity, device fingerprint). Transactions scoring above threshold flagged for Tier 2 analysis. Threshold calibrated to capture 95 percent of actual fraud while flagging only 2.8 percent of all transactions for Tier 2 review. Tier 2 — LLM contextual analysis (new): Llama 3 70B fine-tuned on Korean fraud investigation data, processing flagged transactions with rich context. For each flagged transaction, the LLM analysed: transaction context (the sequence of recent transactions — was this purchase consistent with the cardholder's established pattern? Did the location sequence make geographic sense?), communication context (recent customer service contacts — had the cardholder reported a lost card, asked about unfamiliar transactions, or exhibited behaviour consistent with social engineering?), merchant context (the merchant's transaction patterns — was this merchant showing anomalous transaction patterns suggesting collusion or terminal compromise?), and social network context (graph analysis of connections between cardholders, merchants, and devices — identifying fraud ring patterns). LLM processing time: 1.8 seconds per flagged transaction. Volume: 218,000 flagged transactions daily (2.8 percent of 7.8 million). The LLM providing a fraud probability score and natural-language explanation — enabling fraud analysts to understand why the AI flagged a transaction, not just that it was flagged. Fine-tuning: 18,000 investigated fraud cases with Korean-language analyst annotations — the model learning Korean fraud investigation terminology, Korean social engineering patterns (specific to Korean culture — voice phishing scams targeting elderly Koreans, messenger phishing through KakaoTalk), and Korean merchant fraud typologies. (3) Korean-native customer service AI. Model: Mixtral 8x22B fine-tuned on 52,000 Korean-language customer interactions from the division's contact centre — learning: Korean financial terminology used by actual cardholders (colloquial terms, not formal financial language), appropriate Korean honorific levels (jondaenmal for standard interactions, more formal levels for complaint handling and elderly customers), Korean card-specific concepts (installment payment — halbu — options, point conversion processes, family card management, check card versus credit card differences), and resolution procedures (the division's specific processes for billing disputes, card replacement, limit changes, and fraud reporting). Conversation quality: the AI handling multi-turn conversations in natural Korean — understanding context, following up on previous statements, and maintaining appropriate Korean conversation flow (different from English conversation patterns). Testing: 200 Korean-speaking evaluators (the division's contact centre quality team) blind-testing AI versus human responses — AI responses rated within 5 percent of human quality for standard inquiries. Integration: deployed as a first-contact AI, handling customer inquiries through the division's mobile app (primary channel — 68 percent of contacts), KakaoTalk channel, and web chat. Seamless escalation to human agents for complex issues, with the AI providing a conversation summary to the human agent (in Korean) for continuity. (4) Merchant intelligence platform. Merchant risk scoring: LLM processing merchant data from multiple sources: transaction patterns (volume trends, average transaction size, refund rates, chargeback rates), financial health indicators (from business credit bureau data — Korean business credit reporting), public information (business registration changes, legal proceedings, industry news — processed from Korean-language sources), and network analysis (connections between merchants through shared terminals, common transaction patterns, and ownership structures). Risk model: producing merchant risk scores updated weekly, with high-risk merchants flagged for review by the division's merchant risk team. Early warning: the system identified 34 merchants that subsequently engaged in fraudulent activity (an average of 6 weeks before fraud was detected by traditional monitoring) and 28 merchants that went bankrupt (3 months average early warning — enabling the division to adjust credit limits and reduce exposure). Merchant analytics: AI-powered insights for the division's merchant sales team — identifying: merchant acquisition targets (businesses in high-growth areas with card payment potential), competitive intelligence (merchants accepting competitor cards but not the division's cards — targeting for acquisition), and merchant retention risk (merchants showing declining transaction volume or increasing competitor card acceptance). (5) Korean language fine-tuning: financial Korean mastery. Training corpus: 52,000 customer interactions, 18,000 fraud investigation reports, 8,400 merchant risk assessments, 12,000 internal policy documents, and 4,200 regulatory correspondence documents — all in Korean. Korean financial vocabulary: the model learning 4,800+ Korean financial terms with contextual usage patterns. Korean morphological handling: fine-tuning specifically addressing Korean agglutinative morphology — ensuring the model correctly parsed Korean financial terms when combined with grammatical particles and suffixes. Korean evaluation: native Korean financial professionals evaluating model outputs across: language naturalness (4.6/5.0), financial accuracy (4.4/5.0), cultural appropriateness (4.7/5.0), and honorific correctness (4.8/5.0).

Results over 12 months post-deployment. Fraud detection: overall fraud detection rate from 89.2 percent to 96.8 percent (7.6 percentage point improvement). Annual fraud losses from KRW 82B to KRW 34B (59 percent reduction — KRW 48B in annual savings). False positive rate from 0.42 percent to 0.18 percent (57 percent reduction — 186,000 fewer legitimate transactions incorrectly flagged monthly, dramatically improving customer experience). LLM Tier 2 analysis: 218,000 flagged transactions analysed daily with 1.8-second median processing time. Social engineering fraud detection: 340 percent improvement (from 22 percent to 75 percent detection rate for voice phishing and messenger phishing scams). Merchant fraud early warning: 34 fraudulent merchants identified average 6 weeks early. Customer service: automation rate from 28 percent to 72 percent (of 1.4 million monthly contacts). Customer satisfaction for AI-handled contacts: 88 percent (exceeding the 85 percent target, up from 62 percent with previous rule-based chatbot). Average handle time: from 4 minutes 20 seconds (human) to 1 minute 45 seconds (AI for resolved inquiries). Korean language quality: rated 4.6/5.0 for naturalness by native Korean evaluators. Contact centre efficiency: 180 FTE equivalent capacity freed for complex inquiries and proactive customer engagement. Merchant intelligence: merchant risk early warning: 34 fraud cases (6 weeks early), 28 bankruptcies (3 months early) — preventing estimated KRW 8.4B in exposure losses. Merchant acquisition: AI-identified targets converting at 34 percent (versus 18 percent for traditionally identified targets). Merchant retention: high-risk churn merchants identified and retained at 62 percent rate through proactive intervention. FSC/FSS compliance: annual FSS IT examination: zero findings related to AI systems. Full PIPA compliance for AI data processing. Credit Information Act compliance verified. Korean data residency: all AI processing within Korean-resident infrastructure — no cross-border data transfers. Performance: H100 cluster operating at 64 percent average GPU utilisation (fraud detection peaks during daytime transaction volume). System uptime: 99.97 percent. Fraud detection Tier 2 latency: median 1.8 seconds (within real-time fraud decision window). Cost: infrastructure investment: KRW 5.2B (approximately USD 4M — 32 H100 GPUs, servers, networking). Annual operating cost: KRW 1.4B (USD 1.1M). Annual value: fraud reduction KRW 48B, customer service efficiency KRW 12.6B, merchant risk prevention KRW 8.4B, merchant acquisition improvement KRW 4.2B. Total annual value: KRW 73.2B (approximately USD 56M). ROI: 11.1x in first year.