[{"content":"Overview At Bucketplace (오늘의집), I am building an AI-powered search agent that interprets natural language queries and translates them into structured retrieval signals for Room Planner 3D products. The system decomposes free-form user intent into actionable dimensions \u0026ndash; category, attributes, color, dimensions, and budget \u0026ndash; via LLM reasoning, then executes BM25+KNN hybrid search on ElasticSearch to surface the most relevant 3D product models.\nAt the core of this project is CoI-Fit (Context-Intent Fit Matching), a compositional multimodal retrieval framework I designed to serve as the retrieval backbone for multiple downstream agents. CoI-Fit combines space analysis, mood/style interpretation, dimensional constraints, and conversational context drawn from image, text, and 3D coordinate inputs to produce contextually grounded retrieval results.\nKey Achievements Compositional Retrieval Framework: Designed CoI-Fit, a novel multimodal retrieval architecture that fuses heterogeneous signals (visual, textual, spatial) into a unified retrieval pipeline Two-Tier Negative Query Separation: Hard exclude (must-not) + soft downrank via graduated confidence scoring, preventing irrelevant results while preserving recall 5-Level Progressive Fallback: Graceful degradation from full-signal retrieval down to broad category search, ensuring non-empty results even for ambiguous queries Multi-Agent Architecture: Architected a LangGraph pipeline with parallel fan-out inference, enabling concurrent processing of multiple retrieval dimensions Dual Interface Design: Built both A2A (Agent-to-Agent) JSON-RPC and REST/FastAPI interfaces, enabling seamless integration with both agent ecosystems and traditional service architectures Auto Quality Recovery: Implemented intelligent filter relaxation retry mechanisms that automatically recover from overly restrictive queries, ensuring high recall even for long-tail searches Automated Quality Evaluation: 8-dimension rule-based CI scoring + LLM-as-Judge with LangFuse experiment tracking; persona-based synthetic query generation for long-tail coverage Benchmark Results (634 queries) Metric Improvement Judge Satisfaction +11.9% Category Recall +16.6% Positive Hit +18.0% Negative Leak -50% Latency p50 -20.8% graph LR A[NL Query] --\u003e B[LLM Parse] B --\u003e C[Negative\\nSeparation] C --\u003e D[BM25+KNN\\nHybrid Search] D --\u003e E[Graduated\\nConfidence Scoring] E --\u003e F[5-Level\\nFallback] F --\u003e G[Ranked Results] Technical Approach Pipeline Topology The system follows a multi-stage agentic pipeline with parallel fan-out for inference:\nflowchart TD A[preprocess] --\u003e B1[pre_analyze] A --\u003e B2[pre_category_agg] B1 --\u003e C[planner] B2 --\u003e C C --\u003e D1[query_rewrite] C --\u003e D2[query_embedding] C --\u003e D3[infer_category] C --\u003e D4[infer_attributes] C --\u003e D5[infer_colors] C --\u003e D6[infer_price] C --\u003e D7[infer_dimensions] C --\u003e D8[negative_separation] D1 --\u003e E[retrieve_with_signals] D2 --\u003e E D3 --\u003e E D4 --\u003e E D5 --\u003e E D6 --\u003e E D7 --\u003e E D8 --\u003e E E --\u003e F[twidder] F --\u003e G[confidence_scoring] G --\u003e H[check_refine] H --\u003e|retry\\n5-level fallback| E H --\u003e|final| I[format_response] Request Journey (End-to-End) Intake: Input normalization, safety check, format validation Analyze \u0026amp; Plan: Language/token analysis, category aggregation, search mode determination Parallel Intelligence: Fan-out to 8 concurrent inference nodes — query rewrite, embedding generation, category/attribute/color/price/dimension signal extraction, and negative query separation (hard exclude vs. soft downrank) Retrieval \u0026amp; Ranking: BM25 + KNN hybrid search with two-tier negative filtering and twidder for product_id deduplication Confidence Scoring: Graduated scoring across retrieval signals to rank results by match quality Quality Recovery: 5-level progressive fallback (full-signal → relaxed filters → broad category) with max 1 refine to protect p99 Response: Final items with optional debug metadata (node latency, signals, ES query) CoI-Fit: Compositional Multimodal Retrieval CoI-Fit (Context-Intent Fit Matching) serves as the retrieval backbone for multiple downstream agents:\nSpace Analysis: Understanding room context and spatial arrangement from 3D coordinates Mood/Style Matching: Extracting aesthetic intent from text and image inputs Dimensional Constraints: Filtering by physical size requirements derived from the 3D scene Conversational Context: Maintaining coherent retrieval across multi-turn interactions State Design (3-Layer) Layer Role Properties inputs Request original Immutable artifacts Intermediate outputs Mutable, parallel accumulation outputs Final response Finalized at exit Architecture Integration The system operates as both a standalone service and a domain agent within the AI-AP (AI Agent Platform) orchestrator:\nStandalone: Direct REST/A2A calls for search queries Orchestrated: AI-AP Orchestrator routes search requests via agent capability discovery Clear boundary: Orchestrator handles control plane (routing, fallback, circuit break); search agent handles execution plane (BM25/KNN, signal inference, ranking) Tech Stack Agent Framework: LangGraph, A2A (Agent-to-Agent Protocol), ADK (Agent Development Kit) Observability: LangFuse Search Infrastructure: ElasticSearch (BM25 + KNN hybrid, blue/green index deployment) Embedding Models: SigLIP2, QWEN-3-VL-Embedding-2B (composed text+image embedding) Orchestration: Airflow (batch indexing), K8S Operator Model Serving: Triton Inference Server API: FastAPI, JSON-RPC 2.0 Evaluation: Persona-based synthetic query generation + LLM-as-Judge Future Work CoI-Fit Phase 2: Multi-vector search with intent vector + context vector + preference vector for conversational queries Content Mixing: Blending product results with review/style content for answer-type exploration Personalization: Injecting user behavior features (click/scrap/purchase/dwell) into retrieval/rerank stages Impact This project establishes a foundational retrieval layer for the Room Planner ecosystem at Bucketplace. By serving as the retrieval backbone for multiple downstream agents, CoI-Fit enables a new class of AI-powered interior design experiences where users can describe what they want in natural language \u0026ndash; referencing images, spatial constraints, and stylistic preferences \u0026ndash; and receive precisely matched 3D product recommendations. The dual A2A/REST interface ensures the system integrates cleanly into both the emerging agent-to-agent ecosystem and existing microservice infrastructure.\nPeriod January 2026 - Current | Bucketplace (오늘의집)\n","permalink":"https://jungbaepark.github.io/blog/projects/agentic-3d-model-search/","summary":"Building an agentic multimodal search system (CoI-Fit) for 3D product retrieval at Bucketplace, combining LLM reasoning with BM25+KNN hybrid search across visual, textual, and spatial signals via LangGraph and A2A protocol.","title":"Agentic Compositional Multimodal Natural Language 3D Model Search"},{"content":"Overview At Bucketplace (오늘의집), I built an end-to-end segmentation-to-indexing-to-retrieval pipeline that detects and separates multiple furniture/decor instances from room scene images, connects them to the product index, and achieves performance far beyond naive image-to-image similarity. The project also established the foundation for a 3D object retrieval-based digital twin pipeline.\nKey Achievements DetectRate@10: 2.06% to 28.08% (13.6x improvement over baseline) Lead Time Reduction: 2 weeks to 3 days for segmap refinement pipeline Defective Image Removal: 56.7% defective images removed (more than 2x yield) Catalog Scale: 80K to 440K products (5.5x expansion, filtered from 700K for quality) Category Coverage: 154 to 277 categories (+80%) SAM3D Pipeline: ~1/100 cost vs outsourced Image-to-3D 3D Collaboration: 6,000x cost reduction and 2,200x production increase VLM Refinement Optimization: Switched from QWEN3 8B (1 week+ GPU) to GPT-5-nano with reasoning, completing refinement in 2 days via API parallelization Technical Approach Retrieval Pipeline Architecture The pipeline follows a multi-stage approach:\n[Room Scene] → [Object Detection] → [Query Generation] → [Product DB (ES) Search] → [Post-processing] → [Results]\nv1 to v2 Improvements Component v1 v2 Improvement Product DB Size 80K 440K 5.5x expansion Data Quality Inaccurate filtering 700K → 440K quality filter Removed low-quality products Segmap Quality Low Improved VLM-based verification pipeline Post-processing None Category Boost Category-aware re-ranking VLM Refinement QWEN3 8B (1 week+, 2 GPUs) GPT-5-nano + reasoning (2 days, API parallel) 3.5x faster, higher accuracy Segmap Refinement Pipeline IoM-based Merging: Automatic merging of overlapping instances VLM Category Verification: Filtering incorrectly detected objects via vision-language models Representativeness Scoring: Prioritizing search-suitable segments Experiment Results Best configuration: SAM3 bbox → Full Image Search + Category Boost\nConfiguration DetectRate@10 Hit Rate@10 MRR SAM3 bbox → Full Image + Category 28.08% 33.06% 0.157 OWLv2 bbox → bbox crop + Category 26.56% 30.20% 0.166 SAM3 bbox → segmap crop + Category 26.28% 30.43% 0.152 Baseline (naive SigLIP2) 2.06% 2.06% 0.011 Key findings:\nBbox crop outperforms segmap crop as query (+7.6%p) Full product image indexing outperforms cropped indexing Weak filtering (0.01) outperforms strong filtering (0.05) — excessive filtering excludes correct answers 3D Digital Twin Extension Retrieved products can be further transformed into 3D objects for digital twin placement, collaborated with the XR team for top-N segment 3D reconstruction.\nTech Stack Embedding Model: SigLIP2 Object Detection: SAM3, OWLv2 VLM Refinement: GPT-5-nano (reasoning), QWEN3 8B Instruct 3D Reconstruction: SAM3D Pipeline Search Infrastructure: ElasticSearch (vector KNN) Evaluation: DetectRate@K, Hit Rate@K, MRR Period December 2025 - January 2026 | Bucketplace (오늘의집)\n","permalink":"https://jungbaepark.github.io/blog/projects/digital-twin-scene-retrieval/","summary":"Built an end-to-end scene-to-products retrieval pipeline for Digital Twin at Bucketplace, achieving 13x recall improvement, 6,000x cost reduction in 3D collaboration, and scaling the catalog from 30K to 400K+ products.","title":"Scene-to-Products Retrieval for Digital Twin (Image-to-3D)"},{"content":"Overview OHouseAI is an AI-powered interior design generation service at Bucketplace (오늘의집). I led the systematization of the GenAI workflow, modularizing the architecture into a Pipeline Provider + Subgraph pattern using LangGraph. The service reached Korea #8 ranking in the Graphics/Design category.\nKey Achievements Net Satisfaction Score (NSS): +253% Positive-Negative Ratio (PN): +205% User Engagement: Requests per user increased by 2.3x Retention: Improved from 7% to 14% Latency: Reduced from 71.7s to 35.7s Model Quality: Outperformed GPT-IMAGE-1.5 and Gemini Nano (8.3/10 intent reflection, 8.1/10 context preservation) Technical Approach Pipeline Provider + Subgraph Architecture Modularized the GenAI workflow into composable components using LangGraph:\nPipeline Provider: Centralized orchestration layer managing the lifecycle of generation requests Subgraph Pattern: Each generation capability is implemented as an independent subgraph, enabling independent iteration LLM-as-a-Judge Evaluation Pipeline Built a batch evaluation pipeline for data-driven model selection:\nScale: 4 evaluation rounds x 99 test sets Methodology: Automated side-by-side comparisons with structured scoring: Intent Reflection: How well the output matches the user\u0026rsquo;s design intent Context Preservation: How faithfully the output preserves the original context Application: Used to compare candidate models (including GPT-IMAGE-1.5 and Gemini Nano) for production model selection Monitoring and Observability Production monitoring through LangFuse for end-to-end tracing of the generation pipeline.\nArchitecture Systematization Restructured the existing LangGraph pipeline into a modular Pipeline Provider + Subgraph pattern:\nBefore: Monolithic LangGraph graph handling all generation types in a single flow After: Composable subgraphs per generation capability, orchestrated by a centralized Pipeline Provider with checkpointed state, Redis persistence, and MongoDB storage Benefits: Independent iteration on each generation type, modular prompt management, structured request monitoring, and performance evaluation Tech Stack Agent/Workflow Framework: LangChain, LangGraph Backend: FastAPI, Kafka Storage: MongoDB (domain objects), Redis (LangGraph persistence) Observability: LangFuse Evaluation: LLM-as-a-Judge (4 rounds x 99 test sets) Architecture Pattern: Pipeline Provider + Subgraph Period July 2025 - November 2025 | Bucketplace (오늘의집)\n","permalink":"https://jungbaepark.github.io/blog/projects/ohouseai-genai-workflow/","summary":"Led systematization of OHouseAI\u0026rsquo;s GenAI interior design service at Bucketplace, achieving +253% net satisfaction, 2x latency reduction, and outperforming GPT-IMAGE-1.5 and Gemini Nano. Reached Korea #8 in Graphics/Design.","title":"OHouseAI GenAI Workflow Systematization"},{"content":"Overview At Bucketplace (오늘의집), researched and developed personalized multimodal retrieval systems by combining sequential recommendation models with vision-language models and knowledge distillation techniques. This work explores unifying visual understanding, language comprehension, and user behavior modeling into a single retrieval framework for e-commerce product recommendation.\nKey Achievements Sequential Modeling: Implemented SASRec (Self-Attentive Sequential Recommendation) for user behavior modeling, learning from product page view and click patterns Multimodal Embeddings: Integrated Jina-CLIP-V2 for joint vision-language representations Knowledge Distillation: Applied distillation techniques to create efficient production models from large VLMs Demo Development: Built interactive demo using Streamlit and BentoML Technical Approach Problem Context Existing recommendation systems at Bucketplace (e.g., SASRec-based producers for PDP feed) operate on behavioral signals alone — click patterns, view history, and purchase sequences. These models predict \u0026ldquo;what the user will look at next\u0026rdquo; but lack understanding of why — the visual attributes, style preferences, and aesthetic choices that drive user decisions.\nApproach: VLM-Distilled Sequential Recommendation Teacher Model (VLM): Jina-CLIP-V2 generates rich multimodal embeddings that capture both visual and textual product attributes Student Model (SASRec): A lightweight sequential recommendation model is trained to predict the VLM\u0026rsquo;s embedding outputs, effectively distilling multimodal understanding into an efficient sequential model Metric Learning: PyTorch-Metric-Learning for contrastive training between user interaction sequences and multimodal product representations Efficient Training: QLoRA for parameter-efficient fine-tuning, DeepSpeed for distributed training Integration with Recommendation Pipeline The distilled model slots into the existing Producer → Mixer → Ranker → Twiddler recommendation pipeline at Bucketplace, providing multimodal-aware candidate generation alongside existing behavioral producers (SASRec from PDP click, similar image, popular CTR, etc.).\nTech Stack Deep Learning: PyTorch, HuggingFace, DeepSpeed Training: Lightning AI, QLoRA Metric Learning: PyTorch-Metric-Learning VLM: Jina-CLIP-V2 Sequential Models: SASRec Deployment: BentoML, Streamlit Period May 2025 - June 2025 | Bucketplace (오늘의집)\nImpact This project explores the next generation of product recommendation by bridging the gap between visual understanding and behavioral prediction. By distilling VLM knowledge into efficient sequential models, the approach enables personalized multimodal retrieval that can understand both what users do and what they see, unlocking style-aware and context-aware product discovery at scale.\n","permalink":"https://jungbaepark.github.io/blog/projects/multimodal-retrieval-vlm/","summary":"Developed personalized multimodal retrieval combining SASRec sequential recommendation with Jina-CLIP-V2 vision-language representations, using knowledge distillation for efficient production serving at Bucketplace.","title":"Personalized Multimodal Retrieval with VLM Distillation"},{"content":"Overview At Bucketplace, I delivered measurable OKR impact through production experiments by establishing a reusable experimentation foundation (Query Feature Table, Redis, Server) and shipping end-to-end from PRD/Design Doc through DAG development to production rollout.\nKey Achievements STORE Search Result Page (SRP) \u0026ndash; Deals Ranking Buyer Conversion: +1.99% Click Conversion: +0.83% Special-Offer Exposure: +7.35% Category Product List Page (PLP) \u0026ndash; Price 2.0 Ranking Buyer Conversion: +11.17% Click Conversion: +5.87% Exposure: +27.9% Experimentation Track Record 4 total production experiments conducted 3 winners shipped to production Minimal side effects observed Technical Approach Experimentation Infrastructure The project established a reusable experimentation foundation:\nQuery Feature Table: Computed via Airflow DAGs and served through Redis for low-latency access during search request processing. Feature Serving Pipeline: Query Feature Table flows through Redis into the ranking server. End-to-End Delivery Process Each experiment followed the full delivery lifecycle:\nPRD / Design Doc authorship DAG Development via Airflow Production Rollout Hyperparameter Optimization Developed specialized Grid Search HPO for ranking parameter tuning, with additional tooling from Optuna and Ray Tune. Integrated with the MOHPER framework for multi-objective optimization across CTR, CVR, and exposure metrics.\nTech Stack Feature Serving: Redis, Query Feature Table, Feature-Serving-API Orchestration: Airflow DAGs Search Engine: ElasticSearch HPO: Grid Search, Optuna, Ray Tune, MOHPER Data Processing: PySpark, Athena Period February 2025 - June 2025 | Bucketplace (오늘의집)\n","permalink":"https://jungbaepark.github.io/blog/projects/search-okr-ecommerce/","summary":"Delivered measurable search ranking OKR impact at Bucketplace through 4 production A/B experiments (3 winners shipped), achieving +11.17% buyer conversion on category pages and +7.35% deals exposure on store search.","title":"Search OKR: E-commerce Deals / Price 2.0 Ranking Optimization"},{"content":"Overview At Bucketplace (오늘의집), increased iteration velocity and production tuning coverage by launching tuned parameters on SERP and CPLP, targeting CTR/CVR improvements without harming other metrics.\nKey Achievements Publication: Accepted to CIKM'25 Applied Research Track (DOI: 10.1145/3746252.3761496) Patent: Patent applied Production Impact: Launched tuned parameters \u0026gt;5x on SERP and \u0026gt;3x on CPLP Online A/B Results: SERP CTR +0.95%, Click/Query +1.23%; CPLP CTCVR +1.10%, Purchase/Query +1.62% Domain Expansion: Expanded into the AD domain and adopted as common Search Team stack Technical Approach Framework Architecture MOHPER is a multi-objective hyperparameter optimization framework that automatically tunes ElasticSearch ranking parameters while balancing multiple business metrics simultaneously.\nCore Components Multi-Objective Optimization: Bayesian optimization with Optuna/Ray Tune, jointly optimizing CTR, CVR, and other metrics to avoid \u0026ldquo;whack-a-mole\u0026rdquo; metric trade-offs Transform Functions: Custom conversion functions mapping raw signals to ranking scores, enabling fine-grained control over how features influence search ranking Safety Guardrails: Each optimization run ensures no significant degradation in non-target metrics before promotion to production Algorithm Flow The optimization loop follows:\nSample hyperparameter configuration via Bayesian optimization Construct ElasticSearch QueryDSL with sampled parameters Evaluate against offline metrics (simulated click/purchase) Update surrogate model and propose next configuration Promote best configuration to online A/B test Production Deployment Launched tuned parameters \u0026gt;5x on SERP and \u0026gt;3x on CPLP Each optimization targeted CTR/CVR improvements while ensuring other metrics were not harmed Expanded the framework into the AD domain and integrated it as common stack for the Search Team Nominated for Bucketplace Engineering Awards Tech Stack AutoML: Optuna, Ray Tune Optimization: Bayesian Optimization, Multi-Objective (Pareto) Search: ElasticSearch QueryDSL Configuration: Hydra-core Orchestration: Airflow, Katib Evaluation: Offline simulation + Online A/B testing Period October 2023 - January 2025 | Bucketplace (오늘의집)\nPublications Jungbae Park, Heonseok Jang - \u0026ldquo;MOHPER: Multi-objective Hyperparameter Optimization Framework for E-commerce Retrieval System\u0026rdquo; (CIKM 2025, Applied Research Track; arXiv:2503.05227; DOI: 10.1145/3746252.3761496) Talks CIKM 2025: Oral presentation at the Applied Research Track Bucketplace Internal: \u0026ldquo;AutoML for Retrieval (ElasticSearch) System Tutorials\u0026rdquo; (2024.10.24) ","permalink":"https://jungbaepark.github.io/blog/projects/mohper-automl-ecommerce/","summary":"Built a multi-objective hyperparameter optimization framework for e-commerce search at Bucketplace, launching tuned parameters 5x+ on SERP and 3x+ on CPLP for CTR/CVR gains. Accepted to CIKM'25 Applied Research Track.","title":"Multi-Objective Model Hyperparameter Optimization for E-commerce Search"},{"content":"Overview At Bucketplace (오늘의집), built a two-stage content ranking system for 3 content types with time decay and a global language analyzer, served via Redis caching.\nKey Achievements Built two-stage ranking pipeline (candidate retrieval + function-score reranking) for 3 content types Integrated time decay into ranking Implemented a global language analyzer for search quality Redis-cached serving for low-latency responses graph LR A[Query Input] --\u003e B[Candidate\\nRetrieval] B --\u003e C[Function-Score\\nReranking] C --\u003e D[Time Decay\\n+ Language] D --\u003e E[Redis Cache] E --\u003e F[Ranked Results] Technical Approach Stage 1 — Candidate Retrieval ElasticSearch multi_match queries with a global language analyzer (Korean synonym analyzer) across multiple fields: title, card descriptions, hashtags, user nicknames/companies, style names, area names, and residence information.\nMatch Strategy: best_fields with tie_breaker=0.2 for initial matching Recall Optimization: minimum_should_match: 30% for OR-based search to maximize recall Phrase Matching: Variable slop parameters tuned per field — slop=1 for hashtags, slop=2 for titles, slop=3 for longer descriptions Stage 2 — Function-Score Reranking Reranking candidates using ElasticSearch function_score queries with most_fields to accumulate relevance scores across multiple field matches:\nMatching Features: Multi-match and phrase matching scores Embedding Features: Cross-modal similarity (dense vector matching) Feedback Signals: Click-through data from user interactions Temporal Features: Recency decay and engagement velocity Quality Signals: Content quality scoring and sorting mechanisms Content Type Unification Unified 3 existing content collections (photos, home tours, know-how) under a consolidated \u0026ldquo;Content\u0026rdquo; tab with a common ranking pipeline.\nServing Architecture Data processing via Airflow batch jobs with results cached in Redis for low-latency direct client access, serving both app and web interfaces.\nTech Stack Search Engine: ElasticSearch (multi_match, function_score, phrase matching) Orchestration: Airflow (batch processing) Backend: Go Server Caching: Redis Ranking Signals: BM25, cross-modal embeddings, click feedback, time decay Period July 2023 - September 2023 | Bucketplace (오늘의집)\n","permalink":"https://jungbaepark.github.io/blog/projects/global-content-search-ranking/","summary":"Built a two-stage content ranking system at Bucketplace covering 3 content types, with ElasticSearch function-score reranking, time decay, global language analysis, and Redis-cached serving.","title":"Global Content Search Ranking System"},{"content":"Overview At Bucketplace (오늘의집), shipped multimodal retrieval improvements and quantization optimizations for e-commerce search, along with product classification across a large category space.\nKey Achievements Query CTR: +3.03% improvement Query CTCVR: +16.39% improvement Inference Speed: 2x speed improvement via quantization Disk Usage: 1/4 disk reduction through quantization Classification: Top-10 Accuracy 92.23% among \u0026gt;2,000 product categories Technical Approach Multimodal Dense Retrieval Integrated CLIP-based dense vector retrieval into the e-commerce search pipeline, enabling semantic matching beyond traditional BM25 keyword search:\nImage-to-Text Retrieval: Users can search products using text queries that match against product image embeddings Partial Hybrid Search: Combined dense image retrieval via CLIP with traditional text-based BM25, creating a hybrid retrieval system Query Embedding Service: Built a dedicated service for real-time query vector generation Production Optimization INT8 Quantization: Achieved 2x inference speed and 1/4 disk usage with minimal quality degradation Serving Architecture: Deployed via Triton Inference Server for efficient batch inference Product Classification Built a multi-label classification system across \u0026gt;2,000 product categories:\nTop-10 Accuracy: 92.23% — enabling automated category assignment at scale Application: Used for product catalog enrichment and search filter generation Tech Stack Model: CLIP (OpenAI), SigLIP Optimization: INT8 Quantization, ONNX Runtime Retrieval: Dense Vector Retrieval, Hybrid Search (BM25 + KNN) Serving: Triton Inference Server Search: ElasticSearch Period January 2023 - June 2023 | Bucketplace (오늘의집)\n","permalink":"https://jungbaepark.github.io/blog/projects/clip-ecommerce-multimodal/","summary":"Shipped CLIP-based multimodal retrieval at Bucketplace, achieving +16.39% query CTCVR, 2x inference speed via quantization, and 92.23% Top-10 accuracy across 2,000+ product categories.","title":"CLIP-based E-Commerce Multimodal Search and Classification"},{"content":"Overview At Bucketplace (오늘의집), improved data quality monitoring by implementing client-side log validation and log anomaly detection notifications.\nKey Achievements Designed shared ML feature mart for the Search Team Implemented client-side log validation at the ingestion layer Built lag anomaly detection with automated notifications Created automated A/B test analysis pipeline for experiment reporting graph TD subgraph Sources[Data Sources] S1[Client Logs] ~~~ S2[Server Logs] ~~~ S3[Search Logs] end subgraph Processing P1[Log Validation] ~~~ P2[Anomaly Detection] end subgraph FeatureMart[Feature Mart] F1[Shared ML Features] ~~~ F2[Query Features] end subgraph Experiment E1[Offline Eval] ~~~ E2[A/B Analysis] ~~~ E3[Reporting] end Sources --\u003e Processing --\u003e FeatureMart --\u003e Experiment Technical Approach Search Mart v2 Pipeline Designed and built the Search Team\u0026rsquo;s shared ML feature mart (Search Mart v2), consolidating search logs, server logs, and client interaction logs into a unified feature pipeline for downstream ML models and experiment analysis.\nClient-Side Log Validation Implemented DataQualityOperator for validation at the log ingestion layer, catching schema mismatches, missing fields, and malformed events before they enter the pipeline.\nLog Anomaly Detection Built automated lag anomaly detection with notification alerts, enabling the team to detect and respond to data freshness issues and pipeline failures proactively.\nAutomated A/B Test Analysis Created a standardized experiment analysis pipeline producing automated reports for Search Team A/B tests — reducing manual analysis effort and ensuring consistent metric computation.\nTech Stack Data Processing: PySpark, Athena Data Warehouse: HiveDB Orchestration: Airflow (DataQualityOperator) Monitoring: Anomaly detection, automated alerting Period April 2023 - June 2023 | Bucketplace (오늘의집)\n","permalink":"https://jungbaepark.github.io/blog/projects/shared-team-ml-mart/","summary":"Designed a shared ML feature mart for Bucketplace Search Team with client-side log validation, lag anomaly detection, and automated A/B test analysis pipeline using PySpark and HiveDB.","title":"Design \u0026 Development of a Shared Team ML Mart"},{"content":"Overview At RIIID (뤼이드), developed an end-to-end automatic speech scoring system for second language learners that addresses the cold-start item problem by incorporating prompt awareness. Proposed prompt embeddings combined with question context (BERT/CLIP) to enable reliable scoring even for unseen prompts. The system was successfully deployed to SANTA Say, a TOEIC Speaking practice app.\nKey Achievements Publication: Accepted to InterSpeech 2023 (top conference on speech and audio processing) Production Deployment: Integrated into SANTA Say TOEIC Speaking App Cold-Start Solution: Proposed prompt embeddings + question context (BERT/CLIP) to address the cold-start item problem Real-World Impact: Enabling thousands of language learners to improve their speaking skills Technical Approach Traditional speech scoring systems struggle when encountering new prompts (questions) they have not seen during training. This is the cold-start item problem. Our approach incorporates:\nPrompt Embeddings: Learned representations of question prompts that capture their semantic content Question Context via BERT/CLIP: Leveraging pre-trained language and vision-language models to encode prompt text and images End-to-End Architecture: Joint optimization of prompt understanding and speech scoring components This allows the system to provide accurate scores even for new, unseen prompts by understanding the context of what learners are responding to.\nTech Stack Domain: Speech Scoring, TOEIC Speaking Models: PyTorch, BERT, CLIP Problem: Cold-start Item Problem Deployment: Production-grade serving for mobile app Period October 2022 - January 2023\nImpact Enables fair and consistent scoring for speaking practice, helping language learners track their progress and identify areas for improvement, even as new practice questions are added to the platform.\nPublications Jungbae Park, Seungtaek Choi - \u0026ldquo;Addressing Cold Start Problem for End-to-end Automatic Speech Scoring\u0026rdquo; (InterSpeech 2023, DOI: 10.21437/Interspeech.2023-533) ","permalink":"https://jungbaepark.github.io/blog/projects/speech-scoring-interspeech/","summary":"Developed a prompt-aware automatic speech scoring system at RIIID that solves the cold-start item problem using BERT/CLIP prompt embeddings. Published at InterSpeech 2023 and deployed to SANTA Say TOEIC Speaking app.","title":"Prompt-Aware Speech Scoring System for Second Language Learners"},{"content":"Overview At RIIID (뤼이드), proposed Attentive Conditional Contrastive Learning (ACCL) and Relation Contrastive Learning (RCL), achieving state-of-the-art on student (user) modeling\u0026mdash;including dropout prediction, conditional dropout prediction, and knowledge tracing\u0026mdash;across 6 benchmarks. Deployed to the Santa TOEIC platform.\nKey Achievements SOTA on Student Modeling: Achieved state-of-the-art on dropout prediction, conditional dropout prediction, and knowledge tracing with proposed ACCL across multiple benchmarks Multi-Dataset Validation: Validated approach across 6 different datasets Production Deployment: Deployed to Santa TOEIC app via BentoML Novel Methods: ACCL (trainable attentional coefficients reweighting CL objectives) and RCL (RelationNCE loss) Technical Approach Dropout Prediction Predicting whether a student will churn from the learning platform, enabling early intervention strategies. Covers both naive dropout prediction and conditional dropout prediction (conditioned on contextual information). Achieved SOTA with the proposed ACCL method.\nKnowledge Tracing Predicting whether a student will correctly answer a question based on their historical interaction patterns. Improved performance with contrastive learning approaches.\nSAICL: Auxiliary Interaction-level Contrastive Learning The SAICL framework introduces a novel student modeling approach that addresses the sparsity problem in user interaction sequences by:\nSelf-supervised contrastive objective at each interaction step (not just sequence-level) Auxiliary interaction-level CL that helps the model distinguish between user behavior dynamics across sessions Conditional dropout prediction (CondDP) — a new task formulation beyond standard knowledge tracing SURCL: Sequential User Representations via Relation Contrastive Learning SURCL introduces learnable attentional coefficients for RelationNCE loss that:\nPay attention to relations between patches rather than just point-level contrasts Adaptively mediate contrastive objectives based on conditional input similarity Incorporate both conditional inputs and output embeddings for better user representations Generalize across multiple downstream tasks (knowledge tracing, dropout prediction, sequential recommendation) Tech Stack Framework: PyTorch Architecture: Transformer, Sequential Modeling Method: Contrastive Learning, Knowledge Tracing Deployment: BentoML Period August 2021 - October 2022\nImpact Better understanding of student learning patterns enables personalized learning recommendations, early intervention for struggling students, and improved content difficulty calibration on the Santa TOEIC platform.\nPublications Jungbae Park, Jinyoung Kim, Soonwoo Kwon, Sang Wan Lee - \u0026ldquo;SAICL: Auxiliary Interaction-level Contrastive Learning for Knowledge Tracing and Dropout Prediction\u0026rdquo; (Submitted to AAAI 2023; arXiv:2210.09012) Jungbae Park, Soonwoo Kwon, Jinyoung Kim, Sang Wan Lee - \u0026ldquo;SURCL: Sequential User Representations via Relation Contrastive Learning\u0026rdquo; (Submitted to NeurIPS 2022) ","permalink":"https://jungbaepark.github.io/blog/projects/knowledge-tracing-contrastive/","summary":"Proposed ACCL and RCL contrastive learning methods at RIIID, achieving state-of-the-art on student modeling across 6 benchmarks (dropout prediction, knowledge tracing). Deployed to Santa TOEIC platform.","title":"Knowledge Tracing with Contrastive Learning"},{"content":"Overview At RIIID (뤼이드), built a model registry using MLFlow and dataset pipelines using Airflow, Athena, and BigQuery, serving 4+ products.\nKey Achievements Built model registry with MLFlow Built dataset pipelines with Airflow, Athena, and BigQuery Infrastructure served 4+ products: SANTA TOEIC, IVYGlobal SAT, CASA GRANDE, and INICIE graph TD subgraph Products P1[SANTA TOEIC] ~~~ P2[IVYGlobal SAT] ~~~ P3[CASA GRANDE] ~~~ P4[INICIE] end subgraph Serving S1[BentoML] ~~~ S2[Model Registry] end subgraph Training T1[MLFlow] ~~~ T2[Multi-GPU DDP] end subgraph Pipeline L1[Airflow DAGs] ~~~ L2[Dataset Pipeline] end subgraph Storage D1[Athena] ~~~ D2[BigQuery] ~~~ D3[S3] end Products --\u003e Serving --\u003e Training --\u003e Pipeline --\u003e Storage Technical Approach Model Registry: Built on MLFlow for model versioning and experiment tracking Dataset Pipelines: Orchestrated with Apache Airflow, using AWS Athena and GCP BigQuery for data processing Containerization: Pipeline components containerized with Docker Tech Stack Model Management: MLFlow Orchestration: Apache Airflow Data Processing: AWS Athena, GCP BigQuery Containerization: Docker Products Served: SANTA TOEIC, IVYGlobal SAT, CASA GRANDE, INICIE Period January 2021 - September 2021 | RIIID (뤼이드)\nTalks PyCon KR 2021: \u0026ldquo;하나의 코드 베이스, 파이프라인으로 여러 도메인에 AI 모델들을 배포할 수 있을까\u0026rdquo; (Can We Deploy AI Models Across Multiple Domains with a Single Codebase and Pipeline?) ","permalink":"https://jungbaepark.github.io/blog/projects/ml-infrastructure-riiid/","summary":"Built ML model registry (MLFlow) and dataset pipelines (Airflow, Athena, BigQuery) at RIIID, serving 4+ products including SANTA TOEIC, IVYGlobal SAT, CASA GRANDE, and INICIE.","title":"ML Model Registry, Dataset Pipeline \u0026 Infrastructure at RIIID (뤼이드)"},{"content":"Overview At RIIID (뤼이드), introduced the company\u0026rsquo;s first multi-GPU training and improved all pipeline procedures.\nKey Achievements GPU Utilization: Increased from 25% to 95% Initialization Time: Reduced from 1 hour to 10 seconds First Multi-GPU Training: Introduced the first multi-GPU training in the company CI/CD: Built CI/CD pipelines using GitHub Actions graph LR A[Single GPU\\n25% util] --\u003e B[DDP Setup] B --\u003e C[Multi-GPU\\nTraining] C --\u003e D[Docker\\nContainer] D --\u003e E[Distributed\\n95% util] Technical Approach Introduced multi-GPU training to the company for the first time Improved all pipeline procedures, resulting in GPU utilization going from 25% to 95% and initialization time dropping from 1 hour to 10 seconds Set up CI/CD with GitHub Actions Tech Stack Training: Multi-GPU, Data-Distributed Training (PyTorch DDP) CI/CD: GitHub Actions Infrastructure: Docker, AWS Period June 2020 - December 2020 | RIIID (뤼이드)\n","permalink":"https://jungbaepark.github.io/blog/projects/ml-pipeline-acceleration/","summary":"Introduced RIIID\u0026rsquo;s first multi-GPU training, boosting GPU utilization from 25% to 95% and cutting initialization time from 1 hour to 10 seconds. Built CI/CD pipelines with GitHub Actions.","title":"ML Pipeline Acceleration \u0026 Multi-GPU Training at RIIID (뤼이드)"},{"content":"Overview As Research Lead and COO at Humelo (휴멜로), led the development of next-generation speech synthesis and voice conversion systems. This work produced two ICASSP publications: Duration Controllable TTS (ICASSP 2019, Oral, 1st author) and Emotional Voice Conversion (ICASSP 2020). The project was funded by the Ministry of SMEs and Startups through the TIPS R\u0026amp;D grant.\nKey Achievements ICASSP 2019 (Oral, 1st Author): Duration Controllable TTS with phonemic-level duration control ICASSP 2020: Multi-speaker, multi-domain emotional voice conversion Government Grant: Secured TIPS R\u0026amp;D grant from Ministry of SMEs (S2644149) Awards: Minister of Science and ICT Special Award at K-Startup 2018 Technical Approach Duration Controllable TTS (ICASSP 2019, Oral) Achieved natural speech synthesis through phonemic-level duration control using teacher attention alignment. This approach allows fine-grained control over the timing and rhythm of synthesized speech by leveraging attention alignment information from a teacher model to guide duration prediction.\nEmotional Voice Conversion (ICASSP 2020) Developed multi-speaker, multi-domain emotional voice conversion using Factorized Hierarchical Variational Autoencoder (FHVAE). This architecture disentangles speaker identity, emotional expression, and linguistic content into separate latent representations, enabling flexible control over each factor independently.\nKey contributions:\nSequence-level and segment-level disentanglement using FHVAE to separate speaker identity from emotional expression Emotion embedding with margin loss to further facilitate emotion conversion via cycle-consistency loss Multi-speaker, multi-domain setup: 2 speakers x 6 emotion classes, evaluated with MOS (Mean Opinion Score) surveys Tech Stack Core: E2E TTS, Neural Vocoder, TensorFlow Duration Control: Attention Alignment, Teacher-Student Architecture Voice Conversion: Disentangled Representation, Factorized Hierarchical VAE Domain: Multi-speaker, Multi-domain Emotional Speech Period April 2018 - March 2020\nAwards Impact As a co-founder, this work helped establish Humelo as an innovator in AI-driven audio technology. The dual publications at ICASSP demonstrated both the scientific rigor and practical applicability of the research, covering the full spectrum from speech synthesis to voice conversion.\nPublications Jungbae Park, Kijong Han, Yuneui Jeong, Sang Wan Lee - \u0026ldquo;Phonemic-level Duration Control Using Attention Alignment for Natural Speech Synthesis\u0026rdquo; (ICASSP 2019, Oral, DOI: 10.1109/ICASSP.2019.8683827) Mohamed Elgaar, Jungbae Park, Sang Wan Lee - \u0026ldquo;Multi-speaker and Multi-domain Emotional Voice Conversion Using Factorized Hierarchical Variational Autoencoder\u0026rdquo; (ICASSP 2020, DOI: 10.1109/ICASSP40776.2020.9054534) Hyunmook Park, Jungbae Park, Sang Wan Lee - \u0026ldquo;End-to-end Trainable Self-Attentive Shallow Network for Text-Independent Speaker Verification\u0026rdquo; (arXiv:2008.06146, 2020) Patents Voice conversion system and method (KR 1022772050000) Apparatus for synthesizing speech and method thereof (Multiple patents) ","permalink":"https://jungbaepark.github.io/blog/projects/emotional-tts-humelo/","summary":"Led development of duration-controllable TTS and emotional voice conversion at Humelo, producing two ICASSP publications (2019 Oral 1st author, 2020). Won Minister of Science and ICT Special Award at K-Startup 2018.","title":"Emotional Text-to-Speech and Voice Conversion Systems"},{"content":"Overview At Humelo (휴멜로), I planned, proposed, and led multiple government and public R\u0026amp;D projects, securing approximately US$875K in total funding across three competitive grants from Korean government agencies.\nKey Achievements IITP: Brain-Inspired AI (No. 2019-0-01371) Funding: ~US$225K Period: 2019-2020 Role: Co-PI Scope: Development of brain-inspired AI Ministry of SMEs/Startups TIPS R\u0026amp;D (S2644149) Funding: ~US$400K Period: 2018-2020 Role: Leading researcher; co-manager; co-proposal-author Scope: Next-generation deep-learning emotion/expression TTS Seoul R\u0026amp;BD Program (CY190019) Funding: ~US$250K Period: 2019-2020 Role: Leading researcher; co-manager; co-proposal-author Scope: Voice Conversion (VC) and TTS for VC Technical Approach Grant Proposal and Management For each grant, I contributed to proposal writing, research planning, and project management across Humelo\u0026rsquo;s research and engineering teams.\nTech Stack Research Areas: Brain-inspired AI, Emotional TTS, Voice Conversion Management: Grant proposal writing, budget planning, milestone tracking Awards Period April 2018 - March 2020 | Humelo (휴멜로)\n","permalink":"https://jungbaepark.github.io/blog/projects/government-rd-grants/","summary":"Secured ~US$875K across three competitive Korean government R\u0026amp;D grants (IITP, TIPS, Seoul R\u0026amp;BD) at Humelo, covering brain-inspired AI, emotional TTS, and voice conversion research.","title":"Government \u0026 Public R\u0026D Grant Management"},{"content":"Overview At Humelo, I managed and developed convolutional bidirectional LSTM with synthetic data-based transfer learning for polyphonic sound event detection. Published at ICASSP 2019 as corresponding author.\nKey Achievements F1 Score: +28.4% improvement on TUT 2016 dataset Error Rate: -0.42 reduction on TUT 2016 dataset Publication: Published at ICASSP 2019 (corresponding author) Technical Approach Convolutional Bidirectional LSTM with Transfer Learning Developed convolutional bidirectional LSTM architecture Applied synthetic data-based transfer learning Evaluated on TUT 2016 Sound Event Detection dataset Tech Stack Framework: TensorFlow Method: Transfer Learning, Convolutional Bidirectional LSTM Period June 2018 - May 2019 | Humelo (휴멜로)\nPublications Seokwon Jung, Jungbae Park (corresponding author), Sang Wan Lee - \u0026ldquo;Polyphonic Sound Event Detection Using Convolutional Bidirectional LSTM and Synthetic Data-based Transfer Learning\u0026rdquo; (ICASSP 2019, DOI: 10.1109/ICASSP.2019.8682909) ","permalink":"https://jungbaepark.github.io/blog/projects/sound-event-detection-transfer-learning/","summary":"Developed convolutional bidirectional LSTM with synthetic data-based transfer learning for polyphonic sound event detection at Humelo, achieving +28.4% F1 improvement. Published at ICASSP 2019 as corresponding author.","title":"Polyphonic Sound Event Detection with Transfer Learning"},{"content":"Overview At Humelo (휴멜로), led the development of AI systems for music composition, lyrics generation, and rap synthesis. This project combined deep learning with music theory to create AI-aided music, culminating in collaborations with major K-pop artists, presentation at SXSW 2019, and a feature in a KBS Documentary with professional rapper Sleepy.\nKey Achievements SXSW 2019: Presented AI-aided music production at South by Southwest in Austin, Texas SM Entertainment Collaboration: Worked with SM Entertainment through KOCCA\u0026rsquo;s \u0026ldquo;Music X A.I.\u0026rdquo; program KBS Documentary Feature: Featured on KBS1 Documentary \u0026ldquo;New-contents, changing the world\u0026rdquo; with professional rapper Sleepy Media Coverage: Covered by 10+ national media outlets Artist Collaborations: Created AI-generated music with Korean hip-hop band XXX, rapper Sleepy, and other artists Patents: Multiple registered patents for polyphonic music generation systems Technical Approach ALPHACH - LSTM-based Chord Progression Melody Composing Developed an AI system (ALPHACH) for music composition using LSTM-based models capable of:\nGenerating chord progressions based on music theory constraints Creating melodic compositions over generated chord sequences Polyphonic music generation using LSTM and polynomial approaches Lyrics Generation and Rap Synthesis Created a system for automated lyrics generation and rap synthesis that:\nGenerates contextually coherent rap lyrics Synthesizes rap vocals for collaboration with human artists Featured in the collaboration with professional rapper Sleepy on KBS Documentary AI-Aided Music Production (SXSW 2019) SXSW 2019 Showcase: \u0026ldquo;XXX - Flight Attendant\u0026rdquo;\nIntegrated sound and visual interpretation Collaborated with XXX (Korean Hip-hop Band), Seonggu de Kim (Sound Designer, DJ), and Seong Pil Kim (Visual Artist) Tech Stack Deep Learning: Seq2Seq, LSTM, Polynomial models for music generation NLP: Lyrics generation, text processing Audio: Sound synthesis and rap vocal generation Framework: TensorFlow Period 2017 - 2019\nImpact This work demonstrated that AI could be a creative partner in artistic endeavors, opening new possibilities for human-AI collaboration in music and entertainment. The extensive media coverage across 10+ national outlets and the KBS Documentary feature brought significant public attention to AI-generated music.\nPublications Jungbae Park, Jaryong Lee, Sang Wan Lee - \u0026ldquo;A New Approach for LSTM Polynomial Melody Composition based on Finite Chord Progression\u0026rdquo; (Korea Intelligent Information System Conference, 2017 Spring) Patents System, Device, and Method to Generate Polyphonic Music (KR 1022274150000, registered; PCT/KR2019/003333 applied) ","permalink":"https://jungbaepark.github.io/blog/projects/ai-music-composition-sxsw/","summary":"Led AI music composition and rap synthesis at Humelo, presented at SXSW 2019, collaborated with SM Entertainment and rapper Sleepy (KBS Documentary), and received coverage from 10+ national media outlets.","title":"AI Music Composition and SM Entertainment Collaboration"},{"content":"Overview At Humelo (휴멜로), I managed and developed a speech emotion classification system for the Emotional TTS pipeline. The system performed feature extraction and multi-class emotion detection from audio signals.\nKey Achievements Built feature extraction and classifier for multi-class emotion detection from audio signals Integrated the emotion recognition module into the Emotional TTS pipeline graph LR A[Audio Input] --\u003e B[MFCC + Mel\\nExtraction] B --\u003e C[SpeechCNN / CRNN] C --\u003e D[Emotion\\nClassifier] D --\u003e E[Multi-class\\nOutput] Technical Approach Feature Extraction Implemented feature extraction using MFCC and Mel-spectrogram analysis from audio signals.\nClassifier Architecture Developed and evaluated two architectures:\nSpeechCNN: A convolutional neural network for audio-based emotion classification CRNN: A convolutional recurrent neural network combining CNN with recurrent layers Multi-Class Emotion Detection The system classified audio segments into multiple emotion categories.\nTech Stack Deep Learning Framework: TensorFlow Architectures: SpeechCNN, CRNN Audio Features: MFCC, Mel-spectrogram analysis Domain: Speech emotion recognition, multi-class classification Period April 2018 - February 2019 | Humelo (휴멜로)\n","permalink":"https://jungbaepark.github.io/blog/projects/speech-emotion-recognition/","summary":"Built a multi-class speech emotion recognition system at Humelo using SpeechCNN and CRNN architectures with MFCC/Mel-spectrogram features, integrated into the Emotional TTS pipeline.","title":"Speech Emotion Recognition \u0026 Classification System"},{"content":"Overview Master\u0026rsquo;s thesis research at KAIST exploring attentional control methods for time-series data classification and synthesis. This work investigated how attention mechanisms can improve deep learning models\u0026rsquo; ability to process sequential data, with applications in neuroscience and signal processing.\nKey Achievements Publication: Oral presentation at IEEE SMC 2018 Developed novel attention-based approaches for time-series classification and synthesis Applied reinforcement learning for adaptive EEG signal processing Solved the memory-based vs. memoryless trade-off problem Technical Approach EEG Signal Classification Applied reinforcement learning for adaptive EEG signal processing, addressing the memory-based vs. memoryless trade-off problem. This approach enables the model to dynamically decide when to rely on historical context and when to focus on current observations.\nAttention Mechanisms for Sequential Data Investigated how attention can focus on relevant time steps in sequential data, balancing model capacity with interpretability. Applications span neuroscience and signal processing domains.\nTech Stack Deep Learning: LSTM, Attention Mechanisms Signal Processing: EEG Analysis, Time-Series Processing Frameworks: TensorFlow, PyTorch Period March 2017 - February 2019\nImpact This foundational research in attention mechanisms laid the groundwork for later work in multimodal AI and sequential modeling, including knowledge tracing and recommendation systems.\nPublications Jungbae Park, Sang Wan Lee - \u0026ldquo;Solving the Memory-based-Memoryless Trade-off Problem for EEG Signal Classification\u0026rdquo; (IEEE SMC 2018, Oral) Thesis Title: \u0026ldquo;Attentional Control Methods for Time-series Data Classification and Synthesis\u0026rdquo; Advisor: Prof. Sang Wan Lee Institution: KAIST, Lab for Brain and Machine Intelligence ","permalink":"https://jungbaepark.github.io/blog/projects/attentional-control-timeseries/","summary":"Master\u0026rsquo;s thesis at KAIST on attentional control for time-series classification and synthesis, solving the memory-based vs. memoryless trade-off for EEG signals. Oral presentation at IEEE SMC 2018.","title":"Attentional Control for Time-Series Data (Master's Thesis)"},{"content":"Overview Undergraduate research project at KAIST exploring how artificial agents can learn cognitive policies through competitive multi-agent reinforcement learning. This work was inspired by how humans develop cognitive abilities through social interaction and competition.\nKey Achievements Best Paper Award: 2016 Fall Korea Intelligent Information System Conference Developed multi-agent RL framework for cognitive policy learning Demonstrated emergence of complex behaviors through agent competition Explored connections between computational models and cognitive science Technical Approach Investigated whether AI agents can develop sophisticated cognitive strategies by competing with each other, similar to how humans learn through social competition and cooperation. The framework allows agents to develop increasingly complex policies through iterative competitive interactions.\nTech Stack Method: Multi-Agent Reinforcement Learning Domain: Cognitive Science, Policy Learning Institution: KAIST, Lab for Brain and Machine Intelligence Period March 2016 - February 2017\nImpact This was the first major research project as an undergraduate at KAIST, conducted as part of Individual Research at the Lab for Brain and Machine Intelligence. The insights from multi-agent learning influenced later research on sequential modeling, user behavior understanding, and brain-inspired AI approaches.\nPublications Jungbae Park, Juno Kim, Sang Wan Lee - \u0026ldquo;Multi-agent Cognitive Policy Learning: Reinforcement Learning Through Competition\u0026rdquo; (Korea Intelligent Information System Conference, Fall 2016) - Best Paper Award ","permalink":"https://jungbaepark.github.io/blog/projects/multi-agent-reinforcement-learning/","summary":"Undergraduate research at KAIST on multi-agent cognitive policy learning through competitive reinforcement learning, demonstrating emergence of complex behaviors. Won Best Paper Award at 2016 KIIS Conference.","title":"Multi-Agent Cognitive Policy Learning through Competition"}]