My Projects

Challenge

The primary challenge addressed in this project was the persistent issue of inaccurate Kanji readings and inherent biases in LLMs. These models often default to popular but incorrect readings due to biases in their training data. This problem is critical, as inaccuracies can propagate misinformation and reduce the reliability of language technologies. Additionally, detecting and mitigating these biases while ensuring the models can flag uncertainties posed a significant challenge.

Solution

To tackle these issues, I employed a dual approach: advanced prompt engineering and targeted fine-tuning of the GPT-3.5-turbo model. I designed and tested a variety of prompting techniques to guide the LLMs towards accurate Kanji readings, focusing on both simple and complex multi-component prompts. I also fine-tuned the model on a curated dataset emphasizing correct Kanji pronunciations and uncertainty elements. This approach aimed to recalibrate the model's tendency to default to incorrect readings and enhance its ability to detect and respond to uncertainties.

Key Findings

• Effectiveness of Complex Multi-Component Prompts: Prompts like Few-shot, Chain-of-Thought, Generated-Knowledge, and Tree-of-Thought significantly improved the model's accuracy in reading Kanji characters and mitigating biases.
• Fine-Tuning Enhances Simple Query Prompts: Fine-tuning the model on a dataset focused on correct readings and uncertainties led to substantial performance gains in accuracy and bias mitigation, particularly for simple query prompts.
• Uncertainty Detection Capabilities: While complex prompts improved uncertainty detection, there remains a need for further refinement to ensure models can effectively flag ambiguous or incorrect readings.
• Bias Recognition and Mitigation: The project highlighted the models' limitations in recognizing and expressing uncertainties when influenced by biased training data, underscoring the importance of tailored training processes.
• Cultural and Perspective Biases: The study revealed how different cultural and temporal contexts, as well as the engineers' perspectives, can introduce biases in LLM training, impacting the accuracy and reliability of generated information.

Impact

This project has had a significant impact on the field of natural language processing and education. It provided a new benchmark for the accuracy of Kanji readings in LLMs, offering valuable insights for Japanese language educators on effective prompting techniques. Additionally, it raised critical awareness of the unique challenges posed by Kanji readings and the broader implications of biases in LLMs. The methodologies developed in this project have paved the way for more reliable and responsible language technologies, contributing to ongoing efforts to improve LLMs' accuracy and trustworthiness in handling complex linguistic challenges.

Snapshots

Link

📝Project Report: https://lip.theaisprint.info/report_llm-evaluation_kanji-reading

Challenge

Clients expressed a desire for a seamless way to leverage the power of LLMs within their existing Google Sheets workflows.

Solution

Developed a user-friendly Google Apps Script application that allows users to:

• Execute multiple prompts within Google Sheets, retrieving responses from various LLMs.
• Dynamically integrate generated information into subsequent prompts, facilitating iterative and refined results.

Key Features

• Multi-Model Support: Offers compatibility with several leading LLMs, providing flexibility and choice to users.
• Dynamic Prompt Chaining: Enables the integration of previous prompt outputs into subsequent prompts, allowing for complex and nuanced interactions with LLMs.

Impact

This application has empowered numerous users to streamline their workflows, boost productivity, and unlock new possibilities within Google Sheets. 

Snapshots

Link

🖥️Google Apps Script application: https://lip.theaisprint.info/g-sprint-v20-english
🖥️Demo Video: https://lip.theaisprint.info/gsprint-v20-english-instruction

Challenge

Creating effective instructions and prompts is crucial for developing high-performing LLM agents, but this process can be time-consuming and require specialized expertise.

Solution

Developed an application that automates the generation of tailored instructions and prompts for specific LLM agent tasks based on simple user inputs.

Key Features

• Multi-Step Prompt Workflow: Features a sophisticated system for designing chained prompts, enabling LLM agents to utilize previously generated information for more sophisticated and accurate results.
• Customizable Instruction Flows: Allows users to define the flow and logic of their instructions, enabling the creation of highly customized LLM agents.
  

Impact

This application empowers AI community members to develop advanced LLM agents with greater efficiency, lowering the barrier to entry for sophisticated agent creation.

Snapshots

Link

🖥️Automated Instruction & Prompts Generator: https://agent-instruction-generator.streamlit.app/
🖥️Source code(GitHub): https://github.com/aisprint/agent-instruction-generator

🖥️Demo Video: https://lip.theaisprint.info/agent-instruction-generator-video

Challenge

As part of Lam Research Corporation, a world-leading semiconductor manufacturing equipment supplier, I contributed to the development of cutting-edge products for clients at the forefront of AI chip manufacturing.

Solution

By forging strong relationships with clients and deeply understanding their unique challenges, I played a key role in the successful development and delivery of a groundbreaking product.

Key Contributions

• Needs Assessment: Conducted thorough market research and client interviews to identify unmet needs in the AI chip development process.
• Competitive Analysis: Evaluated competitor products and technologies to identify opportunities for differentiation and innovation.
• Product Roadmap: Collaborated with cross-functional teams to define the product roadmap and ensure alignment with overall business objectives.
• Game-Changing Innovation: Conceived and championed a novel technical solution that directly addressed a critical customer pain point. This innovation proved to be instrumental in securing key client partnerships.
• Product Launch and Support: Played a key role in product launch activities, including beta testing, customer training, and ongoing technical support.

Impact

My contributions resulted in a highly successful product launch, generating over $20 million in revenue and solidifying Lam Research Corporation's position as a leader in the AI chip manufacturing equipment market.