Project Summary: Led Team NEWBIES to second place in the 4th UNIST-KAIST-POSTECH AI & Data Science Competition, developing a sophisticated financial RAG system among 62 teams with 237 participants, earning recognition for technical innovation and practical application potential.
The Financial AI Challenge
Korea's top three science universities challenged teams to build advanced question-answering systems capable of handling complex financial queries using large-scale document corpora. Our mission was to develop a Retrieval-Augmented Generation system that could rival human financial experts in understanding and answering nuanced financial questions.
The Innovation
Hybrid Search Architecture We developed a sophisticated retrieval system combining Voyage-3 embeddings for semantic understanding with BM25 for keyword matching. Our multi-stage filtering and Reciprocal Rank Fusion approach achieved 8-20 percentage point improvements across all financial datasets compared to single-method baselines.
Advanced Reranking Pipeline Our cross-encoder reranking system integrated commercial APIs (Cohere, Voyage) and open-source models (Jina, BGE, NV-ReRankQA) for progressive relevance scoring. This two-stage approach from lightweight to sophisticated models optimized both accuracy and computational efficiency.
LLM Ensemble Strategy We combined DeepSeek-v3 (671B) and fine-tuned f1-Llama3.3-70B with Chain-of-Thought prompting and LLMLingua-2 compression. Our adaptive model selection and confidence-weighted voting achieved exceptional performance across diverse financial benchmarks, with accuracy improvements of 4-9 percentage points for complex reasoning tasks.
The Results
Our system achieved 2nd place and the Gold Prize along with $1,500 in prize money. Performance highlights include 95.6% accuracy on FinanceBench, 92.3% on FinDER, and significant improvements across ConvFinQA, FinQA, and FinQABench datasets.
As team leader and presenter, I coordinated the technical strategy and delivered the final presentation that showcased our innovations to industry judges.
Industry Impact
The project received media attention in E-Today and generated interest from financial technology companies for potential implementation. Our technical approach demonstrates immediate applications for investment research, compliance monitoring, customer service, and risk management in financial services, while contributing to the advancement of retrieval-augmented generation systems for domain-specific applications.