🌾⚡ Rakshak–Artha
The Autonomous Financial Defense & Credit Intelligence Layer for Bharat
💡 Inspiration
India is undergoing a massive digital transformation, with millions of rural users entering the financial ecosystem through UPI. However, this inclusion comes with a critical imbalance. These users are highly vulnerable to digital financial fraud due to low awareness and lack of protection mechanisms, and at the same time, they are excluded from formal credit systems because they lack traditional credit histories such as CIBIL scores.
Rakshak–Artha was built to bridge this gap. Our goal was to design an enterprise-grade, AI-powered financial system that protects users in real time while converting their digital behavior into meaningful credit signals. The system operates at scale, communicates in native languages, and leverages modern Lakehouse architecture to unify data engineering and machine learning.
⚙️ What It Does
Rakshak–Artha is a dual-engine FinTech platform consisting of two core components:
🛡️ Rakshak — Fraud Monitoring Engine
Rakshak intercepts UPI transactions in real time and performs anomaly detection within milliseconds. It evaluates transaction patterns, behavioral deviations, and contextual signals to determine whether a transaction should be allowed, reviewed, or blocked.
When a suspicious transaction is detected, Rakshak not only prevents potential loss but also triggers an AI-driven explanation system that communicates the reason for the block to the user in their native language. These messages are aligned with RBI guidelines, helping users understand risks and build financial awareness.
💰 Artha — Micro-Credit Intelligence Engine
Artha analyzes a user’s transaction history to generate an alternative credit profile. Instead of relying on traditional credit scores, it evaluates behavioral financial patterns such as transaction frequency, inflow-outflow consistency, spending habits, and anomaly indicators.
Based on these signals, Artha produces a creditworthiness score and recommends micro-loan limits, enabling financial institutions to extend credit to users who were previously invisible to the system.
🛠️ How We Built It
We implemented Rakshak–Artha using a fully decoupled Lakehouse architecture on Databricks, enabling scalable data processing, model training, and deployment.
🔄 Data Engineering
We used PySpark to ingest, clean, and transform raw transaction data. The pipeline follows a Bronze–Silver–Gold architecture:
- Bronze: Raw transaction logs
- Silver: Cleaned and enriched datasets
- Gold: Feature-engineered datasets for ML
We designed separate pipelines for fraud detection and credit scoring to avoid feature leakage and ensure domain-specific optimization. A dedicated gold_user_credit_features table was created for behavioral aggregation.
🤖 Machine Learning Systems
Fraud Detection (Hybrid Model)
The fraud detection system combines XGBoost and a Deep Q-Network (DQN):
- XGBoost estimates the probability of fraud based on tabular features.
- The DQN learns optimal decision policies (allow, review, block) by maximizing long-term reward.
The policy update follows the Bellman equation:
$$ Q(s, a) \leftarrow Q(s, a) + \alpha \left[r + \gamma \max_{a'} Q(s', a') - Q(s, a)\right] $$
This allows the system to optimize decisions based on real-world costs such as financial loss, user friction, and review overhead.
Credit Scoring Model
The credit engine uses an interpretable gradient boosting model designed for explainability and regulatory compliance. It ensures that decisions can be audited and understood, which is critical in financial systems.
📦 Model Lifecycle & Deployment
All models are tracked and versioned using MLflow. They are deployed as independent REST APIs on Databricks Serverless Compute, allowing scalable, low-latency inference.
🧠 Sovereign AI (RAG Pipeline)
We integrated a multilingual AI layer using the Sarvam-1 (7B) model. To make responses context-aware and reliable, we built a Retrieval-Augmented Generation (RAG) pipeline:
- FAISS for vector search
- Multilingual MiniLM embeddings
- RBI document knowledge base
This enables the system to generate accurate, policy-aligned explanations in multiple Indian languages, improving user trust and comprehension.
🌐 Frontend
The platform is delivered through a Streamlit-based multi-tab application hosted via Databricks Apps. It provides a unified interface for fraud alerts, credit insights, and AI explanations.
🚧 Challenges We Faced
LLM Resource Constraints
Deploying a 7B parameter model in a constrained environment was challenging. We used 8-bit quantization and optimized loading strategies to reduce memory usage from approximately 14GB to around 3.5GB without sacrificing performance.
Lack of Structured Credit Data
The dataset did not include user-level credit labels. To overcome this, we synthesized a multi-tenant user base and engineered behavioral features using PySpark, ensuring realistic transaction patterns.
Feature Leakage Risks
Fraud and credit systems require strict separation of data flows. We designed isolated pipelines and ensured time-aware feature generation to prevent leakage and maintain model integrity.
🧠 What We Learned
We learned that building scalable AI systems is not just about choosing powerful models, but about designing robust data pipelines and efficient architectures.
We also optimized system memory usage significantly:
- Fraud model: ~50MB
- Credit model: ~10MB
- Embeddings: ~470MB
- LLM (quantized): ~3.5GB
This demonstrated that high-performance systems can be built efficiently with the right engineering decisions.
🚀 What’s Next
We plan to extend Rakshak–Artha with voice-based interaction, allowing users to report fraud or dispute transactions using speech in multiple regional languages.
Additionally, we aim to integrate the credit engine with microfinance institutions to enable real-time loan approvals and disbursements, further enhancing financial inclusion.
🌍 Final Vision
Rakshak–Artha is more than a fraud detection system. It is a comprehensive financial intelligence platform designed to:
- Protect users from fraud in real time
- Build trust through transparent AI explanations
- Enable access to credit using behavioral data
- Support users in their native languages
Our mission is to ensure that every digital transaction not only remains secure but also contributes to financial growth and opportunity.
⚡ Closing Statement
Rakshak–Artha transforms digital transactions into trust, and trust into financial empowerment—bringing protection, inclusion, and opportunity to every user.
Log in or sign up for Devpost to join the conversation.