Rishi DevData Scientist | AI/ML Engineer

Applied Data Finance

Behavioral Credit Payment Modeling

A machine learning system for predicting customer payment behavior using engineered behavioral and transaction features.

problem

Anticipate whether a customer will make an upcoming payment, using behavioral and transaction signals rather than static application data.

approach

Extensive feature engineering over behavioral and transaction history, gradient-boosted modeling with XGBoost, class-imbalance handling, and a production-oriented evaluation workflow.

data / architecture

Transaction and behavioral tables queried from Redshift, cleaned and aggregated into a modeling dataset in Python.

Shiv Nadar University

Multilingual RAG-based Legal Assistant

A retrieval-augmented legal assistant for querying Indian Penal Code information using semantic retrieval and a fine-tuned language model.

problem

Legal text is dense and hard to query in plain language, especially across multiple languages and through voice.

approach

Chunk legal documents, embed them with Hugging Face embeddings, retrieve with FAISS, and generate grounded answers with Mistral 7B — available in English, Hindi and Tamil, through text and voice.

data / architecture

Indian Penal Code documents, chunked and indexed as vector embeddings.

MCP Server Ecosystem for Sales Automation

A multi-phase agentic AI workflow for lead prioritization, qualification, proposal generation, and follow-up automation.

problem

Sales workflows are fragmented across CRM, documents and communication tools, with context lost between each step.

approach

A phased agent pipeline where context is passed between stages, backed by MCP servers for CRM, document, communication and enrichment capabilities, with Pub/Sub for real-time notification.

data / architecture

CRM records and enrichment sources exposed to agents through MCP servers.

Agiliz Tech

Financial Document Analysis & Forecasting

An AI-powered financial document analysis and forecasting system combining structured data extraction, time-series forecasting, and conversational querying.

problem

Cost data sits inside financial documents and is slow to extract, forecast and interrogate manually.

approach

Structured extraction from financial documents, ARIMA time-series forecasting on the resulting series, and a GPT-based conversational interface over the results.

data / architecture

Financial documents parsed into structured cost time series.

Customer Segmentation & Analysis

A data science project analyzing Portuguese bank marketing data to identify customer segments and subscription patterns.

problem

Understand which customer segments respond to marketing campaigns and what distinguishes subscribers.

approach

Exploratory analysis, missing-data handling, encoding and scaling, dimensionality reduction with PCA, then K-Means and agglomerative clustering to derive interpretable segments.

data / architecture

Portuguese bank marketing dataset, preprocessed and feature-engineered before clustering.