CPA | AI Practitioner
As a Certified Public Accountant and AI Practitioner, I bridge the crucial gap between traditional finance and transformative artificial intelligence. My mission is to empower businesses by revolutionizing accounting and operational workflows with intelligent automation, unlocking new levels of efficiency and insight.
Licensed CPA with a proven track record in financial reporting, auditing, and tax strategy across diverse industries. Successfully led teams to implement best practices and ensure robust regulatory compliance.
Passionate former accounting educator dedicated to demystifying complex financial concepts and fostering professional growth. Mentored numerous aspiring CPAs, guiding them to certification success.
Championed digital transformation within traditional accounting environments, spearheading initiatives that integrated automation and modernized workflows, resulting in efficiency gains exceeding 40%.
My evolution into AI is driven by a commitment to pioneering innovation within professional services. I focus on developing and implementing AI solutions that address real-world financial challenges and create tangible value.
My journey into fine-tuning LLMs locally for a Philippine tax assistant on an M3 Mac (16GB RAM) was a significant learning curve. With Gemma 2B, I navigated Hugging Face authentication, CMake build complexities, and library conflicts. A `requires_grad` error pushed me to disable gradient_checkpointing, initially causing NaNs. I managed to stabilize training by adopting a very low learning rate, a small rank, and switching to bfloat16. While deployment via Ollama was successful, Gemma struggled to fully overwrite some pre-trained factual errors (like tax rates), likely due to weak fine-tuning signals limited by my hardware and a dataset of only 600+ input/output pairs. Switching to TinyLlama 1.1B, the initial fine-tuning likely resulted in catastrophic forgetting, as it struggled with specific tax facts and showed weaker reasoning. This highlighted that pure fine-tuning faces significant hurdles with factual accuracy on small models, datasets, and constrained memory. My next step is a RAG setup with Llama 3 8B on a dedicated GPU for more robust and reliable results.
Developed a tax assistant using Retrieval Augmented Generation (RAG) with a quantized Llama 3 8B model. This project involved leveraging LangChain for the pipeline, Sentence-Transformers for embeddings, and FAISS-CPU for the vector database. Key learnings included managing software dependencies, the critical importance of context for LLM accuracy, and troubleshooting minor output inconsistencies. This hands-on experience significantly enhanced my understanding of RAG and LLM fine-tuning, paving the way for future explorations with Vertex AI.
Leveraging GPT-4 mini-high as my AI assistant, I created six Python scripts to automate accounting tasks that previously consumed a minimum of 5 hours of manual work. These scripts have drastically cut processing time to under 30 minutes, freeing up significant time for continued learning and innovation. Debugging was the most challenging phase, but with AI assistance, issues that could have taken hours were resolved much faster. I'm now focused on developing AI workflows and agents to fully automate most office tasks by year-end, aiming for substantial boosts in efficiency and cost reduction. My experience confirms that failing to embrace AI isn’t just missing out on innovation; it’s choosing to stand still while the world moves forward.
To deepen my understanding of the models I work with, I've been creating visualizations to explore the inner workings of transformer architectures, often referencing explanations like those provided by Anthropic for Claude. My process involves breaking down the model step-by-step: Input text is first tokenized, with each token converted into a vector via token embeddings to capture contextual meaning. Positional embeddings are then added, assigning a unique position code to each token to help the model understand word order. These combined vectors flow into the self-attention mechanism, where each token assesses all others using query, key, and value vectors to gather rich contextual information and understand inter-word relationships. Following self-attention, each token's representation is refined through a feed-forward neural network (FFNN), which expands, transforms, and then compresses the vector. Critical components like residual connections and layer normalization are applied throughout to stabilize learning and improve performance. Visualizing these components helps solidify my grasp of how transformers comprehend language structure and meaning, making them so effective for tasks like translation, summarization, and text generation.
I am eager to explore AI's potential in finance, collaborate on groundbreaking projects, or assist your organization in navigating the future of accounting. Let's connect and discuss the possibilities.