AI Agent Development Company in Seoul

Electronics Manufacturing / Quality Control / Seoul

A Seoul-based electronics component manufacturer (KRW 3.2T annual revenue, 8,600 employees, 4 factories in Korea and Vietnam, supplying display panels and semiconductor packaging to Samsung, LG, and Apple) needed to automate its quality control operations. The company produced 48M units monthly across product lines. (1) Inspection volume — quality inspection generated 2.8M inspection data points daily (automated optical inspection — AOI, X-ray inspection, electrical testing, and manual visual inspection). Each inspection point: pass, fail, or marginal (requiring human judgment). Current process: automated inspection systems flagged potential defects — quality engineers reviewed flagged units and made disposition decisions (pass, rework, scrap). Daily flagged units: 142,000 (5.1% of production). Quality engineer review: 8 seconds average per flagged unit — but with 142,000 daily flags, the 45-person quality team was overwhelmed. (2) False positive burden — automated inspection systems had a 72% false positive rate (industry typical for high-sensitivity settings). Of 142,000 daily flags: 102,000 were false positives (no actual defect — the unit was good), 34,000 were true defects requiring action, and 6,000 were marginal (requiring expert judgment). Quality engineers spent most of their time reviewing units that were actually fine — creating frustration and fatigue that increased the risk of missing genuine defects. (3) Defect pattern detection — the company experienced recurring defect patterns that were detected too late. By the time a quality engineer noticed a trend (e.g., increasing defect rate from a specific production line), the production run had already produced thousands of defective units. Root cause analysis was reactive — triggered by customer complaints or quality holds, not by proactive pattern detection. (4) Customer quality requirements — Samsung and Apple had strict quality requirements with escalation timelines. Quality holds (stopping shipment when defects are detected) needed to be communicated to customers within 2 hours. 8D reports (root cause analysis and corrective action) required within 48 hours. The manual process: quality hold notification averaged 6 hours (quality engineer discovered issue, informed manager, manager escalated through gyeoljae to quality director, quality director notified customer). 8D report completion: 5 business days average (Samsung requirement: 48 hours). Late notifications and 8D reports had resulted in supplier scorecard demerits and risk of programme loss.

We built the quality control AI agent system over 12 weeks: (1) Defect classification agent: AI agent processing all 2.8M daily inspection data points. The agent: analyses inspection images and data from AOI, X-ray, and electrical test systems, classifies each flagged unit as: true defect (with defect type classification — 28 defect categories specific to the company's products), false positive (no actual defect — unit can proceed), or marginal (requires human expert review), and provides confidence scores for each classification. Training: the agent was trained on 8.4M historical inspection images annotated by senior quality engineers — representing 3 years of inspection data with verified outcomes (units that were dispositioned by engineers and subsequently confirmed through downstream testing or customer feedback). Classification performance: false positive reduction from 72% to 18% — the agent correctly identified 76% of false positives without human review. True defect detection: 99.2% sensitivity (missed defect rate: 0.8% — comparable to the best human inspectors). Daily human review required: from 142,000 flagged units to 38,000 (agent-confirmed false positives removed from queue + marginal cases requiring expert judgment). Quality engineer workload: reduced 73% for routine inspection review. (2) Pattern detection agent: autonomous agent monitoring defect trends in real-time. The agent: analyses defect data across all production lines continuously (not batch — real-time streaming analysis), detects statistical process deviations — identifying when: defect rate exceeds control limits for a specific product/line/shift, a new defect type appears or existing defect type frequency changes, correlation between defect patterns and production parameters (temperature, humidity, equipment settings, material batch), and defect rate differs significantly between shifts or operators (reported at aggregate level — not individual operator identification, per company policy). Alert generation: when a significant pattern is detected, the agent generates: a real-time alert via KakaoWork to the relevant quality team, a preliminary analysis (what changed, when, possible causes based on correlated production parameters), and recommended actions (equipment inspection, parameter adjustment, material batch hold). Alert speed: from 6+ hours (manual detection) to 12 minutes (real-time pattern detection and KakaoWork alert). (3) Customer notification agent: autonomous agent managing customer quality communications within gyeoljae constraints. Quality hold notification: when a defect issue requiring customer notification is detected, the agent: generates a quality hold notification with: affected lot numbers, defect description, preliminary containment actions, and estimated timeline for root cause analysis, routes the notification through the company's gyeoljae approval system (quality engineer approval, quality manager approval, quality director approval — each approver receives a KakaoWork notification with one-tap approve), and upon final approval, sends the notification to the customer through the appropriate channel (Samsung SupplyOn, Apple Supplier Connect). Notification time: from 6 hours (manual process with phone calls and emails between approval levels) to 1.5 hours (AI-prepared notification, accelerated gyeoljae with mobile approvals). 8D report assistance: when a quality hold is issued, the agent: initiates the 8D workflow, pre-populates D1-D3 (team formation, problem description, containment actions) from available data, provides D4 root cause analysis suggestions based on historical similar defects and correlated production parameters, and tracks D4-D8 completion with deadline alerts via KakaoWork. 8D completion: from 5 business days to 1.8 business days (within Samsung's 48-hour requirement). (4) PIPA compliance: the quality control agents process manufacturing data (inspection images, defect data, production parameters) — not personal data in most cases. Where production data correlates with individual operators: aggregate reporting only (per company policy and quality team agreement). Shift-level analysis: defect patterns analysed at shift level, not individual operator level. PIPA privacy impact assessment: not required for manufacturing process data without personal information processing. Data localisation: all data stored in Korean data centres (AWS Seoul — ap-northeast-2).

Quality metrics: false positive reduction from 72% to 18% — quality engineers reviewing 38,000 units daily instead of 142,000 (73% workload reduction for routine review). True defect detection: 99.2% sensitivity maintained — no decrease in defect capture rate. Quality engineer productivity: freed from routine false positive review, engineers focus on: marginal case judgment (expert decisions that require human experience), root cause analysis (investigating defect patterns identified by the agent), and process improvement (developing corrective actions and preventive measures). Defect pattern detection: detection time from 6+ hours to 12 minutes — enabling the quality team to respond to emerging defect patterns before they affect large production volumes. One significant detection: the pattern agent identified a solder paste degradation issue 4 hours into a production shift. Previous detection method (manual trend review at end of shift) would have allowed 12 additional hours of defective production. Early detection: 28,000 units saved from potential defect (estimated value: KRW 840M / approximately $630K for that single detection event). Customer quality performance: quality hold notification from 6 hours to 1.5 hours (75% reduction) — now well within Samsung's 2-hour requirement. 8D completion from 5 business days to 1.8 days — within 48-hour requirement for all customers. Samsung supplier scorecard: quality metrics improved from "Conditional" to "Preferred" status. Apple supplier audit: zero critical findings related to quality response time (from 2 findings in prior audit). Revenue protection: Samsung "Preferred" status enabled the company to bid on a new programme worth KRW 180B annual revenue (previously excluded due to "Conditional" rating). Apple audit improvement: maintained the company's position on a KRW 95B programme that was at risk due to quality response concerns. Operational efficiency: quality engineering team maintained at 45 people despite 15% production volume increase — previously projected to need 58 quality engineers at the increased volume (13 avoided hires at KRW 65M fully-loaded cost = KRW 845M annual savings). Quality team satisfaction: satisfaction survey improved from 3.2 to 4.1/5 — engineers valued the transition from routine review to expert analysis and process improvement. Defect cost reduction: manufacturing scrap rate reduced 22% (early pattern detection preventing defective production runs). Scrap cost savings: KRW 3.8B annually. Cost: quality department efficiency: KRW 845M annual savings (avoided hiring). Scrap reduction: KRW 3.8B annually. Revenue protection/enablement: KRW 275B in programmes secured or enabled. Customer penalty avoidance: KRW 420M annually (late notification and 8D penalties eliminated). Development cost: KRW 280M (approximately $210,000). Payback period: 4 weeks.