ASEP Mid Semester Review Presentation

Applied Science and Engineering Mid Semester Review Presentation

1. Citizen Helpline for Government Schemes and Services DIV: A GROUP: 12 DATE: 26 st April 2025 PROJECT GUIDE: JAYASHRI BAGADE PROJECT TEAM: 67-Tanushri Kankrej - (12415205) 68-Hiral Kawediya- (12415230) 69-Arhum Shah- (12415400) 70-Kshitij Bhosale- (12415362) 71-Harsh Chaudhari- (12415361) 72-Suhani Mahajan- (12415371)

2. PROBLEM IDENTIFICATION: Citizens, especially those in rural and underserved areas, often face difficulties accessing information about government schemes due to: Language Barriers: Limited availability of information in regional languages. Digital Literacy Gap: Inability to navigate text-based digital platforms. Complexity of Information: Difficulty understanding technical or bureaucratic terms used in scheme descriptions. Inefficiency in Communication : Lack of personalized, user-friendly, and readily accessible platforms for querying scheme details. Limited Reach of Existing Systems: Traditional helplines and websites often fail to provide timely and accurate responses. DOMAIN: Public Welfare

3. Literature Review R ef no. P roblem Solution R esult Limitations 1 Citizens struggle with outdated information, limited translation, and no chatbot system, hindering eligibility checks and participation in government schemes. The CCB is a bilingual chatbot on a user-friendly website, using NLP techniques like SVM, TF-IDF, and regular expressions, with voice-to-text and text-to-speech features. The chatbot improves accessibility with dual-language support, simplifies scheme retrieval and eligibility checks, and boosts participation with a seamless voice interface. The system lacks real-time updates, struggles with complex queries, and has limited multi-language support, currently only in English and Tamil. 2 Many struggle to navigate financial assistance programs, as existing platforms lack personalized guidance, leading to confusion and missed opportunities. A chatbot centralizing government scheme info, using NLP, GPT models, and Langchain for dynamic data and personalized recommendations. Available on Android and Web with OAuth authentication. SchemeSetu boosts financial literacy and inclusivity by simplifying access to government schemes, eligibility checks, and user engagement. The chatbot can improve in handling complex queries and providing more personalized recommendations based on individual financial circumstances. 3 Citizens face difficulty accessing government schemes due to complex criteria, scattered information, and time-consuming traditional methods, especially for those with limited digital literacy. An AI-based chatbot using NLP enables citizens to check their eligibility for government schemes through an interactive, user-friendly interface accessible via web browsers. The chatbot streamlines eligibility assessments, providing real-time, accurate responses and improving citizen access to government programs. The system needs regular updates to stay current and may struggle with complex queries. Expanding to mobile platforms requires additional resources.

4. Literature Review R ef no. P roblem Solution R esult Limitations 4 No citizen service channel available for people looking for medicare and help in agriculture. D evelopment of AI chatbot that can resolve queries related to medicare and agriculture. This helped citizens get their queries solved and get benefits of these schemes. This was just limited to schemes like medicare and agriculture. It was also text-based, which made it less interactive and couldn’t reach wider audience. 5 Irresponsive and hard to use government websites. Singapore, South America and states like Telangana developing standalone apps and AI systems for government services. It got really easy for citizens to interact with government services and schemes, and they could get more done in less time. It is difficult for citizens (especially from rural regions) to use app and text-based interfaces. They are more suited to natural interfaces like telephone line. 6 Limited awareness and accessibility of government schemes among rural and remote beneficiaries. Use of mobile applications, SMS-based notifications, and USSD services for scheme updates and awareness. Improved dissemination of information, with beneficiaries receiving timely updates about schemes and services Dependence on mobile penetration; challenges with digital literacy and infrastructure in remote areas.

5. Literature Review R ef no. P roblem Solution R esult Limitations 7 Citizens face delays and inefficiencies due to the manual handling of queries and repetitive tasks. Reduced query resolution time by 50%, increased system efficiency, and improved citizen satisfaction. Improved response rate by 45% and reduced operational costs by 30%. Chatbots struggled with understanding complex or context-sensitive queries requiring human escalation. 8 Inconsistent responses from human agents due to varying knowledge levels and outdated information. Deployment of an AI-driven knowledge base system with machine learning to provide agents with real-time, standardized answers. Accuracy of responses increased by 50%, reducing follow-up complaints. Requires regular updates to the AI model; initial data collection for training was resource-intensive. 9 Difficulty in addressing diverse citizen needs due to language and dialect variations across regions.   Use of Natural Language Processing (NLP) for multilingual support, including regional dialects, with voice and text options. Citizen engagement increased by 35%, especially in rural and regional areas. Limited NLP accuracy for uncommon dialects; periodic updates required to adapt to language nuances.

6. Literature Review R ef no. P roblem Solution R esult Limitations 10 Challenges in delivering personalized learning, managing administrative tasks, and ensuring equitable access to quality education. AI-powered tools provide adaptive learning, automate routine tasks, and support multilingual solutions, improving accessibility and personalization. Improved learning outcomes, reduced administrative burdens for teachers, and increased access to education for underserved communities. Concerns over data privacy, algorithmic bias, and inequitable deployment across socio-economic contexts. 11 Traditional chatbots struggle with handling complex queries, multilingual support, and context-aware responses, leading to citizen frustration. Generative AI improves chatbots with natural language understanding, real-time assistance, and RAG for accurate, context-specific answers, supplemented by human oversight for sensitive issues. Improved citizen satisfaction, enhanced efficiency in query handling, and seamless integration with existing systems. Challenges in ensuring transparency, ethical concerns, and reliance on robust data infrastructure and expertise. 12 Governments face inefficiencies, high costs, and limited citizen engagement due to manual processes and lack of automation. AI automates routine tasks, supports data-driven policymaking, and introduces chatbots for 24/7 citizen assistance, improving service reliability and reducing delays. Increased efficiency, lower operational costs, better citizen experiences, and enhanced policymaking through AI-driven insights. Concerns over data privacy, lack of AI expertise, and biases in algorithms impacting service delivery.

7. Literature Review R ef no. P roblem Solution R esult Limitations 13 Traditional tools struggle with semantic understanding in querying unstructured documents ZenDB integrates LLMs to process SQL-like queries, improving semantic interpretation and query execution. ZenDB demonstrates significant improvements in accuracy and efficiency when compared to traditional approaches. Effectively processes complex queries with reduced computational cost, offering a scalable solution for document analytics. Relies on pre-trained LLMs, may have biases, and faces challenges with real-time applications and domain-specific data. 14 The study investigates the effectiveness of fine-tuned large language models (LLMs) compared to bespoke, custom-trained models in classifying complex legal documents. It compares a bespoke classification model to a fine-tuned GPT-3.5 LLM, assessing their performance on binary and multi-label tasks using 29,307 labeled documents. The bespoke model outperformed LLMs, but fine-tuning improved LLM performance significantly, approaching but not surpassing bespoke model accuracy. Fine-tuning requires less labeled data but cannot match bespoke model accuracy due to inherent dataset noise and diminishing returns with additional fine-tuning. 15 Indonesia's Ministry of Finance struggles with complex and dynamic financial data and regulations, hindering efficient decision-making. KemenkeuGPT, an AI model built with LangChain , RAG, and fine-tuning, provides automated insights from financial data and regulations to support policy decisions Achieved significant performance improvements with 44% correctness, 73% faithfulness, and positive feedback from experts on its decision-making utility. Limited data scope, manual updates, reliance on human feedback for accuracy, and a basic user interface.

8. Literature Review R ef no. P roblem Solution R esult Limitations 16. Elderly individuals in urban resettlement areas may lack awareness and utilization of social welfare schemes designed for their benefit. Conducting a community-based cross-sectional study to assess the awareness and utilization levels among elderly residents. Out of 931 participants, 86.9% were aware of at least one social welfare scheme, and 42.2% utilized them. Utilization was higher among females and those aged 75 and above. The study was limited to a specific urban area in Delhi, which may not be representative of other regions. Self-reported data may also introduce bias. 17. Call centers face challenges in maintaining service quality and efficiency due to high call volumes and variability in agent performance. The study tested the effectiveness of voice-based AI assistants in assisting human agents during customer interactions, aiming to improve productivity and customer satisfaction. The AI assistant significantly enhanced agent performance, especially for less experienced agents, by reducing handling times and improving resolution rates. The study's findings may not generalize across industries or customer demographics, and the AI's long-term impacts on customer satisfaction and agent job roles remain unclear. 18. Lack of efficient methods for integrating voice-based interaction with intelligent systems to assist users in providing evidence-based responses A digital assistant leveraging voice recognition and AI-driven prompting to guide users in generating structured and evidence-backed inputs. The system improves user efficiency and accuracy in producing evidence-based responses compared to traditional methods. Limited adaptability to domain-specific nuances and challenges in handling complex or ambiguous queries.

9. Context of a problem: Knowledge of and access to government schemes and services is scanty. By developing interactive and responsive interfaces to access these schemes and services can result into higher participation from citizens Gap Analysis: Most of the AI helplines aim at text-based and graphics-based interfaces for citizens to engage, citizens with not much experience of these interfaces find it difficult to use them. Most solutions focus on niches like medicare , agriculture, etc., whereas solutions can be developed for wider domain to benefit large population.

10. Problem Statement: Citizens, especially in underserved areas, face significant challenges in accessing information about government schemes due to language barriers, low digital literacy, and complex bureaucratic details. Existing systems often fail to provide user friendly, multilingual, and accurate information. This leads to a lack of awareness and underutilization of benefits. An AI voice-bot is needed to bridge these gaps and ensure inclusivity. Objectives : Prepare Government Schemes Dataset Study and Analyse LLM models. Implement Speech-to-Text and Text-to-Speech using python libraries frameworks like PyTorch and TensorFlow for model development and optimization Performance evaluation of Voicebot. Study and analyse the existing application

11. Technical Approach Speech-to-Text Convert spoken queries into text for processing. Language Model Understand the intent and context of user queries- GEMINI FLASH Frameworks Machine learning frameworks like Pytorch & tensorflow . Text-to-Speech Generate spoken responses for users. Our team will utilize Python for developing the citizen helpline. Python's versatility and extensive libraries make it ideal for building AI-powered applications.

12. Methodology Python Environment Gemini Flash Training Dataset API keys Speech Synthesis Text-to Voice

13. Progress Made So Far

14. 6-Week Plan 1 Project Planning and Research Selecting of required tools and technologies and preparing training dataset. 2 Design and Architecture Importing pre-trained language models into development environment 3 LLM Development Fine-tuning the pre-trained model on training dataset

15. 6-Week Plan 4 Bot Integration Integrating the bot with text-to-speech and voice-to-text interfaces 5 Testing and Feedback Debugging errors and bugs in the developed project 6 Refinement and Deployment Refining and deploying for public use

16. Thank-you We appreciate your time and attention. We would love to hear a feedback for improvements and guidance about how to proceed further.