From Complexity to Clarity: AI’s Role in Narcotics Investigations

In a post-lunch session at the Machine Learning Developers Summit (MLDS) 2024, Siva Prasad Polepally, the Delivery Head of Data & Analytics at Quadrant Technologies, began his presentation by acknowledging the unique challenge faced by the narcotics enforcement department. Handling narcotics cases presented a distinct set of hurdles, especially at the ground level, where officers lacked specific training and readily available information.

Inefficiencies in Handling Narcotics Cases

The Telangana state government approached Quadrant Technologies with a significant issue – inefficiencies in handling narcotics cases. Officers faced challenges in collecting and managing data related to narcotics cases, leading to weak legal proceedings and lost cases. The lack of structured information, varying procedures across regions, and the uniqueness of each case made it difficult for ground-level personnel to handle these scenarios effectively.

The Solution: Introducing the “Nor Guide Bot”

To address this challenge, Quadrant Technologies proposed the development of an AI-powered solution – the “Nor Guide Bot.” This bot aimed to act as an AI companion for narcotics investigations, providing guidance and instructions to ground-level police personnel. The idea was to create a tool that could understand natural language queries from officers and provide accurate, contextually relevant responses.

The Data Journey: Overcoming Challenges in Data Collection

The journey began with challenges in data collection. Two primary sources were identified – the Narcotic Control Bureau website and internal notes on the NDPS (Narcotic Drugs and Psychotropic Substances) Act. However, the data presented challenges due to varying procedures, department-specific information, and documents in different formats. The team at Quadrant Technologies tackled this by creating structured databases, and extracting insights from police reports, case judgments, and cross-examination data.

Data Extraction: Enhancing Accuracy with Table Transformer Architecture

Data extraction, a common challenge with PDF documents, saw the team utilizing OCR (Optical Character Recognition) techniques. However, challenges arose when dealing with tables in documents, leading to accuracy concerns. The introduction of the Table Transformer architecture, focusing on table detection and recognition, significantly improved accuracy, especially in scenarios with complex table structures.

Navigating Challenges in Model Training

The team opted for the GPT-3.5-based model, Lama 65b, for its language capabilities. Challenges emerged during the training phase, particularly in handling hallucination and ensuring context relevance. Ground-level officers might phrase questions differently, leading to varying responses. The team addressed this by refining temperature settings and incorporating context vector databases to enhance relevance and accuracy.

Results and Manual Testing: A Balancing Act

Manual testing with police personnel involved asking 1500 prompts, of which 1243 were deemed correct. The bot achieved an 83.6% accuracy rate, proving effective in guiding officers through processes rather than providing strict legal advice. The team acknowledged the importance of covering edge cases and continuing to refine the bot’s performance.

Reinforcement Learning and Language Adaptability

Looking ahead, Quadrant Technologies aims to incorporate reinforcement learning to enhance the bot’s capabilities further. Additionally, efforts are underway to adapt the bot for Indic languages, considering the diverse linguistic landscape of the user base. Future data collection initiatives include generating RFPs and exploring different forms of structured data.

Conclusion

Siva Prasad Polepally concluded the presentation by showcasing the bot’s user interface and expressing the team’s commitment to ongoing refinement. The Nor Guide Bot represents a transformative solution, leveraging AI to empower ground-level officers in handling narcotics cases efficiently. As Quadrant Technologies continues to explore enhancements and adaptations, the journey reflects the potential of AI in addressing complex real-world challenges.