Applying Large Language Models to Financial Risk Control at Akulaku

Akulaku is an overseas internet finance platform offering services such as online shopping, installment payments, cash loans, and insurance, and it applies AI to financial risk control, e‑commerce intelligent customer service, and recommendation. Manual rule‑based processes cannot efficiently handle large volumes of requests, so the goal is to build an agile, intelligent risk‑control system.

The business scenarios involve multiple data modalities: image data for KYC face verification, text data for intelligent customer service, voice data for voice‑based service and collection, and device data for environment verification and unique ID construction.

The overall LLM deployment strategy consists of three core modules: (1) Planning – encoding business decision knowledge in LangChain chains; (2) Memory – storing various data and metadata in external databases; (3) Tools – domain‑specific models such as image, NLP, and risk‑assessment models. LLM agents augment the existing risk‑control and model systems.

Two reinforcement directions are pursued: (1) Tools reinforcement – improving specific models (e.g., KYC face verification, NLP for customer service) by using LLMs for data augmentation, prompt generation, and distilling knowledge into smaller, faster models; (2) Planning and Memory reinforcement – constructing copilot agents for fraud investigation and data analysis, leveraging GraphRAG for intent recognition and vector stores for document retrieval.

Case study 1 (NLP model optimization) shows that using LLM agents to generate prompts and synthetic data reduces the required labeled samples from tens of thousands to a few hundred, cuts labeling effort by 90%, shortens delivery time, and improves model performance by about 20%.

Case study 2 (image anti‑fraud model) demonstrates the use of CLIP and a vision transformer with three loss functions—contrastive CLIP loss, reconstruction loss, and classification loss—to align visual and textual features, improving generalization for KYC face‑spoof detection and enabling faster model iteration.

Case study 3 (Planning/Memory reinforcement) includes a fraud‑investigation copilot that uses GraphRAG for intent recognition and structured queries, and a data‑analysis assistant (ChatBI) that employs Text2SQL and a Pandas‑based visualization tool, with iterative LLM‑driven SQL generation, validation, and execution.

The summary and outlook outline a two‑step rollout: first, enhance specific models using LLMs to improve efficiency; second, progressively extract and solidify business knowledge into agent chains, creating an AGI‑like system that augments human capabilities rather than replacing them, allowing staff to focus on core business issues and complex cases.

The presentation concludes with thanks to the audience.