Building a Companized Private LLM

In the financial industry, where the accuracy of information is critical, general-purpose LLMs available in the market cannot fully address the challenges that arise. We provided comprehensive support for a company-specific Companized Private LLM — from concept design through build. To minimize risk, we adopted approaches that either built a lightweight model from scratch or fine-tuned a base model. Using foundations such as the Financial-General 7B model, we trained the system on the client's proprietary data and industry expertise, establishing an AI environment capable of generating highly factual and accurate responses. Beyond simple system deployment, we also fed the know-how gained through multiple in-house LLM build projects back into the client organization to drive lasting adoption — supporting AI literacy across employees and the advancement of differentiated, competitive use cases grounded in proprietary data.

[Challenges]

• General-purpose LLMs in the market do not match the client's specific use cases (they require adjustment for each scenario and do not satisfy in-house use cases that demand accuracy).
• The client wished to build a dedicated LLM but lacked the know-how (few internal experts on LLMs, limiting the ability to make confident decisions on whether to proceed).
• Highly differentiated use cases were difficult to identify (use of ChatGPT and similar tools is unlikely to translate into firm-specific differentiation).

[Results]

• Built an LLM with high factual accuracy specialized for the client's industry and operations (steering the system to incorporate the factuality of proprietary data and industry-specific knowledge to ensure accuracy).
• Drew on experience building multiple in-house LLMs to transfer that know-how to the client (using the build process itself to contribute to AI literacy improvement).
• Drove differentiated use cases informed by an understanding of the build process and the data used (with a proprietary LLM, the system can be applied flexibly across business needs and use cases).