AI-Powered Legal Assistance for Bangladeshi Laws using Machine Learning and Large Language Models

Abstract

In Bangladesh, getting legal help is very difficult because the legal system too much complex and most people cannot afford legal advice. The situation is even worse for people in rural areas, since professional legal services are almost never available outside of the big cities. To solve this problem, we created a system called AI Lawyer, which is powered by Artificial Intelligence (AI). The system gives people free, simple and always available legal advice based on Bangladeshi laws. We used Artificial Intelligence (AI) and Natural Language Processing (NLP) so that the system can answer legal questions in Bangla, English and Banglish. To build the dataset, we made a Python scraper that collected legal texts from the official Bangladesh law portal. This data was cleaned, structured and stored in JSON format so that the AI model could understand it properly. To make the system faster and more affordable, we fine-tuned an LLaMA 3.1 8B model using LoRA adapters and 4-bit quantization. For deployment, we used a dual inference system. A locally fine-tuned model running on Ollama gives quick answers, while the Gemini API is used as a backup to ensure continuous service if the local model fails. The chatbot is available through a web-based interface developed with Next.js, React, Node.js and PostgreSQL. We tested AI Lawyer

with real legal questions and found that it provides fast, accurate and easy-to- understand answers. We also compared it with other systems like DoNotPay and

LawBot and found that our solution works much better for the Bangladeshi context. User feedback confirmed that the chatbot is simple to use, dependable and very useful. The AI Lawyer project reduces people’s dependency on expensive lawyers, protects citizens from legal fraud and misinformation and helps them understand their rights more clearly. It makes legal information cheaper, more accessible and available in multiple languages. In the future, we plan to expand the dataset, improve accuracy and add voice-based features so that even more people, including those who cannot read or type easily, will be able to use the system.