Carlo Pinto

Built a comprehensive automation system for Philippine Bureau of Internal Revenue (BIR) compliance workflows. The system automatically processes DAT file submissions, searches for BIR confirmation emails, extracts HTML content, converts to PDF with proper naming conventions, and uploads to Google Drive folders. Overcame challenges including Gmail thread splitting, PDF rendering fidelity, and cached file duplication issues. Implemented cross-reference validation by TIN and file type (QAP, SLSP, SAWT) to ensure data integrity. This automation reduced manual processing from hours to minutes, handling hundreds of client submissions seamlessly.

Developed a privacy-focused AI HR agent for resume-job description matching, running entirely on-premises. Built with Dockerized Ollama REST API using quantized Llama 3 (8B), integrated via LangChain in Flask. Implemented PyPDFLoader for resume parsing, SentenceTransformers for embeddings, and FAISS-CPU as vector store. The system generates AI-driven candidate fit evaluations while maintaining strict data confidentiality - crucial for our firm's requirements. Currently enhancing with quantitative match scoring and exploring more capable language models while prioritizing the balance between performance and privacy.

Created multiple Python automation scripts that transformed our tax and accounting workflows. Developed a GUI app combining quarterly income tax processing, cutting time by 70%. Built intelligent file renaming systems that match TINs against client databases and append taxpayer names automatically. Implemented context engineering techniques with AI assistants, completing complex automation tasks in 13 minutes that would've taken full days manually. These tools demonstrate how AI-assisted development accelerates traditional accounting processes while maintaining accuracy and compliance.

Pioneered local LLM fine-tuning for Philippine tax assistance on M3 Mac (16GB RAM). Started with Gemma 2B and TinyLlama 1.1B, navigating technical challenges including gradient checkpointing, NaN losses, and catastrophic forgetting with limited hardware. Evolved to implement RAG architecture using quantized Llama 3 8B, LangChain pipeline, and FAISS vector database. This progression from pure fine-tuning to RAG demonstrated critical lessons about factual accuracy limitations in constrained environments and the importance of retrieval-augmented approaches for domain-specific applications.

Currently developing a comprehensive AI-powered accounting and tax platform built on Tryton ERP, designed for deployment to paid clients. Integrating OCR and LLM capabilities with custom modules for Philippine withholding tax and mandatory government contributions (SSS, PhilHealth, Pag-IBIG). Also exploring Odoo Community Edition with custom HR Payroll modules tailored to Philippine regulations. This project leverages my extensive ERP experience from Nestlé, Maersk, and SC Johnson to create practical, scalable solutions that bridge traditional accounting with AI innovation.

Continuously evaluate emerging AI models for practical accounting applications. Tested GPT-OSS 20B/120B (MoE architecture with 5.1B active parameters), GLM 4.5 (hybrid inference modes for agent applications), Kimi K2 (1 trillion parameters Chinese model), Horizon Beta (256K context, 150 tokens/sec), Grok 4 Heavy (multi-agent system), and SmolLM3-3B (WebGPU browser deployment). Also experimented with ChatGPT Agent for autonomous calendar management and tested various models for complex accounting problems, architectural modeling, and code generation. This ongoing research ensures we leverage the most capable models for specific use cases.

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: tokenization, embeddings, positional encoding, self-attention mechanisms, and feed-forward networks. Critical components like residual connections and layer normalization are visualized to understand how transformers comprehend language structure and meaning, making them effective for translation, summarization, and text generation.