Apr 2024 – Jun 2024 · 2 mos · On-site

• Reconciled accounts by comparing the final consistency of asynchronous databases. Used Hive for offline queries to reduce the load on the online database. Addressed accounting errors caused by long message chains in the Finance Shared Service Center.
• Used large language models (LLMs) to extract contract content and stored enumerated type variables in a vector database. Reduced token usage by 20% by employing vector similarity to match the LLM's answers with actual enumerated types, avoiding the need to include all hundreds of enum values in the prompt. Enabled dynamic addition and modification of enum objects.
• Developed a tool to test the accuracy of LLM-extracted content: Used SQL to query the database and obtain contract information and object storage addresses of contract texts. Retrieved contract texts and used LLM plugins and skill prompts to extract key contract information. Automatically determined whether each contract field was required (e.g., start and end dates for fixed-term contracts). Compared results with manual annotations and calculated accuracy as the number of correctly extracted fields divided by the total number of fields that needed extraction by the LLM.

Jun 2023 – Sep 2023 · 3 mos · On-site

• API Integration Expertise: Developed the trading system with a focus on API integration with the Singapore Exchange's OMNet system, enhancing connectivity and data exchange.
• Adapter Layer Implementation: Engineered a robust Adapter layer to encapsulate low-level details, delivering a standardized interface for the strategy layer. This abstraction facilitated more efficient strategy deployment and adaptability.
• Market Data Delivery: Provided real-time, deep market data and detailed contract specifics, crucial for enabling advanced trading strategies. This initiative significantly improved low-lantency decision-making processes and trading outcomes.
• Unified Factory for Streamlined Operations: Introduced a unified factory concept, streamlining object creation and scheduling. Enabled the strategy layer to uniformly process information from diverse exchanges, fostering a more integrated and effective trading approach.
• Optimized Market Data Management: Innovated market data handling by first using the query command to establish a baseline of market conditions, creating an in-memory market data database. Subsequently, by subscribing to broadcasts of price changes above this baseline, the market data database was maintained. This approach not only reduced latency but also conserved network bandwidth, enhancing overall system efficiency and performance.

Mar 2023 – Jun 2023 · 3 mos · On-site

• Development of DIOPI (Device-Independent Operator Interface): Created the Device-Independent Operator Interface, DIOPI, facilitating seamless interaction between advanced deep learning frameworks and chipset computational libraries. This initiative successfully decoupled the framework's upper layers from the chipset's lower layers, streamlining the implementation process.
• Architectural Design and Implementation: Conceptualized and developed DIOPI-PROTO, a pivotal runtime standard prototype interface for DIOPI, including a broad range of operators, establishing a robust framework for future development and integration.
• Specialized Implementation Layer (DIOPI-IMPL): Engineered a specialized implementation layer for the specific chipset to implement operators in their own way. Adapting NVIDIA's CUDA operators into the DIOPI framework, guaranteed the adapter layer transparent to CUDA users.
• Operator Registration and Automated Code Generation: Innovated a system for operator registration, leading to the streamlined automated generation of code. This enhancement significantly improved efficiency and reduced the potential for manual errors in code creation.
• Model Validation and Integration: Demonstrated the framework's versatility and robustness by successfully integrating various models, thereby validating the effectiveness and reliability of the DIOPI framework in a real-world context.
• Troubleshooting and Continuous Operation: Identified and rectified erroneous operators, implementing strategies to bypass such obstacles, ensuring uninterrupted operation throughout processing epochs. Compiled a comprehensive list of unimplemented operators, setting the stage for future enhancements and development.
• This role underscored my ability to navigate complex technical challenges, drive innovation, and deliver solutions that enhance operational efficiency and adaptability in the rapidly evolving field of deep learning technology.