CHEERLA NITHINSAI

Data Scientist

Hyderabad, India 3+ yrs exp 86 · Excellent

About

Data Scientist spanning the full ML lifecycle in financial services — from explainable credit-risk modeling to production deployment and applied GenAI. Builds transparent, regulation-aware risk models, productionizes them as automated cloud pipelines, and develops RAG/LLM tools that streamline underwriting workflows.

Skills & Expertise (41)

Python Advanced

8.5/10

Years Exp

Machine Learning Advanced

8.2/10

Years Exp

Gradient Boosting Advanced

8.0/10

Years Exp

Predictive Modeling Advanced

8.0/10

Years Exp

Feature Engineering Advanced

8.0/10

Years Exp

SQL Advanced

8.0/10

Years Exp

Deep Learning Intermediate

7.5/10

Years Exp

AWS Intermediate

7.5/10

Years Exp

stakeholder communication Intermediate

7.0/10

Years Exp

Data Storytelling Intermediate

7.0/10

Years Exp

scikit-learn Conversational AI Business Acumen Pandas NumPy Streamlit MySql Git DVC Power BI Tableau Matplotlib Seaborn Cross-Functional Collaboration Classification Feature Selection Statistical Modeling Generative AI NLP RAG PySpark Supervised Learning Unsupervised Learning Prompt Engineering Model Evaluation validation Hyperparameter Tuning Model Interpretability Class Imbalance Handling LLMs Vector Databases

Work Experience

Data Scientist

Kapitus

Feb 2025 - Present

Credit Risk Model — New & Renewal (v5.0–v5.1) | Python, LightGBM, ShapRFECV, AWS • Lifted risk-ranking performance ∼9% (AUC 0.73 vs. 0.67 prior) — sharpening decision precision and reducing losses — by developing a LightGBM classification model to predict customer risk. • Expanded predictive coverage with 4,400+ features engineered from transaction and statement data and 8,000+ external attributes from multiple third-party data providers. • Reduced the feature space from 8,800+ to 392 high-signal predictors using SHAP-based RFE (ShapRFECV), applying encoding and transformations to improve robustness and interpretability. • Enabled explainable decisioning in production by supporting deployment on a hosted model-serving platform (Abacus.AI) that returns probability scores, risk tiers, and SHAP-based reason codes. • Delivered an early ∼4% performance lift on v5.1 through richer behavioral features and expanded data coverage. Renewal Eligibility (Behavior) Model — Early-Delinquency Scoring | Python, CatBoost, SHAP/RFE, SQL • Developed a CatBoost model on refreshed customer + third-party data to flag high-risk accounts before renewal, achieving AUC 0.85 on a held-out future window (0.90 in-sample, KS 54.6), powering an automated eligibility-decisioning process. • Ran a SHAP/RFE feature-selection sweep across 1,200+ behavioral attributes, narrowing to 77 high-signal features (∼94% smaller), and defined the prediction target (6-month default, ∼75% of bad accounts). • Validated on an out-of-time window with decile rank-ordering and ROC/KS, benchmarking risk tiers against prior models to confirm consistency and incremental lift. • Operationalized monthly scoring of the eligible portfolio to flag high-risk renewals earlier, reducing exposure on recent cohorts. ClearReport — RAG-Based Underwriting Document Chatbot | RAG, AbacusAI • Created and deployed a RAG chatbot on AbacusAI that lets users query 50+ page documents, cutting processing from 55 minutes to under 5 minutes (∼91% reduction). • Engineered document classification (>95% accuracy), section-aware information extraction, and LLM summarization, driving an 85% reduction in manual errors. • Delivered conversational Q&A with context retention and parallel processing — 200+ documents/hour, up to 50 concurrent users, and <3s response time. Behavior Score for Equipment Finance — Rule-Based Scoring | Python, SQL, AWS • Designed and shipped a Python & SQL rule-based scoring engine grading ∼955K records on a 0–10 risk scale from engineered delinquency, write-off, legal, and account-status signals. • Constructed a weighted scoring framework that prioritizes recent delinquency over historical performance, with critical-flag overrides and a new-customer carve-out, producing transparent, reason-coded scores for business users. • Streamlined end-to-end monthly batch scoring with point-in-time data-quality controls across 50K+ active accounts, publishing risk-segmented outputs to an AWS S3-backed SFTP endpoint for a new product line.

Research Experience

IIT Bhubaneswar

Aug 2022 - Dec 2023

Application of IMDAA Datasets in Drought Analysis • Analyzed 30 years of precipitation data in Govindpur Watershed, Odisha, to characterize drought conditions and regional vulnerabilities. • Cut manual processing time by 80% by automating SPI calculations in MATLAB, enabling real-time drought assessment.