As the fast evolution of generalist LFMs is quickly reaching a plateau and towards the general AI platforms, the next phase of AI research is on "AI for X", the adaptation of LFM platforms to perform discipline-based research in specific verticals. The research focus then moves from building the platform itself to extending the platform to solve high-stakes vertical domain problems. The platforms here include both state-of-the-art closed and open-source systems. E-commerce has been covered under the recommendation research discussed above, while Healthcare represents a highly promising frontier where we are actively exploring LFM applications in clinical decision support, patient communication, medical documentation, and long-horizon care workflows. In the following, we will briefly outline our research focuses on finance here.
-
AI for Finance: At the very high-level, we can broadly divide the key players in finance domains into 3 categories, which include: (A) large banks and national level financial institutions, (B) mid-sized financial institutions and various loan-based corporations, and (C) quantitative finance companies. The entities in Category A are heavily legislated with key customers in government and huge corporations. They have less appetite for AI-based automation other than RPA (Robotic Process Automation) that involves automating various heterogenous information processing to extract key information and signals to assist human analysts in a wide range of applications like customers onboarding, loan processing, credit assessment, etc. Entities in category B are more aggressive on RAP range of AI applications and services. Category C entities are advanced users of AI for various investment and decision supports. However, their requirements are specific, specialized and less scalable. AI technologies are used to make application level, operational-level and strategic level decisions.
-
Our research focusses on developing RPA-based processes to support a range of financial services, and assist application and operation level decisions. They can be summarized into 5 key research directions. The first is the development of safe and trustable Finance-LLMs capable of processing numeric, tabular and time-series data that are dominant in finance domain. The second is in designing Financial Agents capable of timely and accurate retrieval, synthesis and analysis of multi-modal and heterogeneous financial data. The third is in ensuring compliance with financial rules and regulations by service Agents. This is achieved by incorporating knowledge, requirements and regulations into reasoning process, with robust guardrail mechanisms to guarantee full compliance. The fourth is in developing a comprehensive and rigorous evaluation framework. Finally, we plan to consolidate our research towards the concept of AI-enabled banks to support a wide range of financial services, with initial focus on servicing Category-B financial entities.