## Pilot Service: AI-Powered Proxy Voting Recommendations
What is This Service?
This pilot platform generates proxy voting recommendations using insights derived from my co-authored paper: Beyond Bias: AI as a Proxy Advisor. It leverages an AI model with its knowledge domain of proxy voting, offering shareholders, governance professionals, and researchers a complementary and experimental approach to traditional proxy advisory recommendations.
Why This Matters
Transparency: Each recommendation is generated using the standardized prompt structure outlined in the paper. The versions for the AI model and other tools are all disclosed in this website.
Innovation: Leverages the capabilities of AI (Language Model) to interpret the context of shareholder proposals with relevant knowledge domain (Proxy Voting).
Accessibility: Provides an academic lens on ESG issues for anyone with an interest in corporate governance and shareholder activism.
How It Works
Proposal Input: Enter a shareholder proposal of your choice, with your supporting arguments for it, via Google Form here.
Research Matching: An internal system identifies all the relevant input components for the AI model's final prompt. For instance, the system retrieves and processes historical opposition statements used against similar proposals. These statements are hand-collected from SEC proxy filings and prepared for the referenced academic paper.
Recommendation Generation: The AI model produces a recommendation (e.g., For / Against).
Given the limited time and resources available:
Proposals are accepted throughout the week, and all submission from the previous week are processed on Monday. Reports will be sent to the email address provided in the Google Form.
For each weekly batch, only the most recent proposal submitted by the same requester will be considered.
Current Model & Tools:
AI Model: Llama 3.3 70B Instruct model (commit 6f6073b423013f6a7d4d9f39144961bfbfbc386b) with bfloat16 for the model precision
Tools: Transformers library (Hugging Face) version 4.55.0, with PyTorch 2.5.1 + CUDA 12.4 in Python 3.12.7
Differences between the Referenced Academic Paper and This Service
The Paper: Reproducibility of results is essential for academic evidence. Therefore, I deliberately restrict the AI model from deviating, requiring it to select the output token with the highest probability within its knowledge domain.
This Service: In contrast, this pilot service is intended as a practical and exploratory experiment. Here, I allow the model greater flexibility and creativity, which means that even for the same proposal, the generated recommendations may vary.
Disclaimer for This Pilot Service
This is an early-stage pilot service, and I welcome constructive feedback. choonsiklee79@gmail.com
This service is operated solely with my own time, effort, and resources. Processing times may vary depending on the availability.
The reports generated by this service do not constitute tax, legal, insurance, or investment advice, nor do they represent a recommendation, offer, or solicitation to buy or sell any product, vehicle, service, or instrument. By using this service, you acknowledge that you have read and understood this disclaimer and agree to release me from any liability related to its use.
Version History
2025-09-03: Pilot Service Started