Available for opportunities · 2026

Nellimarla Lalith Chaitanya

Data Scientist & ML Engineer with real experience in churn prediction, fraud detection, RAG systems, and computer vision research — not just notebooks.

0.85
F1 Score · VICTOR
60×
Latency Reduction
16.8×
ROI on Churn System
DATA
SCIENCE
📊
🤖
🧠
🔍
📈
🛠️
🎯
01

About

I'm a Data Science & AI undergrad at IIIT Kottayam, currently interning as a Computer Vision & Data Science Researcher at NIT Trichy. I build end-to-end systems — from raw, messy data through feature engineering, modeling, explainability, and production deployment.

My work spans applied domains including finance, customer churn, lifetime value modeling, fraud detection, and recommendation systems. I care about getting evaluation right: no leakage, honest metrics, business-grounded framing.

I compete on Kaggle and participate in hackathons to keep my edge sharp. Outside the ML stack, I'm working toward a publication on my multi-camera violence detection research — currently under peer review.

💳
Finance & Fraud
Behavioral signals, anomaly detection, imbalanced classification
📉
Churn & LTV
Forward-looking labeling, SHAP explainability, A/B decision engines
🎯
Recommendations
Collaborative filtering, cold-start, hybrid retrieval systems
🤖
LLM & RAG
Hybrid retrieval, LangGraph agents, FAISS, prompt injection defense
📡
Time-Series
SCADA forecasting, LSTM pipelines, R² ≈ 0.97 on industrial data
🎥
Computer Vision
Re-ID, spatiotemporal modeling, CPU-optimized real-time inference
02

Experience

Jun 2025 — Dec 2025
NIT Trichy
Trichy, Tamil Nadu
NIT Trichy
Computer Vision & Data Science Research Intern
  • Developed VICTOR — a real-time multi-camera violence detection framework with identity-aware tracking, cross-camera person Re-ID, and aggressor attribution.
  • Built a feature extraction pipeline over RWF-2000, generating 5,292 temporal sequences with 541-dimensional multimodal behavioral representations.
  • Designed a lightweight BiLSTM temporal model achieving F1-score 0.8513 and 91.25% recall for violence detection.
  • Reduced model complexity to 169K parameters, delivering 60× lower latency and 6× faster inference than 3D CNN baselines on CPU.
  • Performed extensive ablation and runtime scalability studies, validating deployment across increasing numbers of tracked identities.
📄 Research paper "VICTOR: Identity-Centric Temporal Modeling for Multi-Camera Violence Detection and Aggressor Attribution" — currently under peer review.
03

Projects

01
User Retention & Churn Prediction System

End-to-end customer intelligence platform on 541K+ transactions. ETL pipelines, feature store, churn/LTV modeling, SHAP explainability, and a real-time decision engine with A/B testing — identifying £1.89M revenue at risk.

XGBoostLightGBMSHAP FastAPILangGraphFAISSStreamlit
AUC 0.82 R² 0.99 16.87× ROI £1.89M risk flagged
2026
GitHub →
02
RepoRAG — Production RAG over Live GitHub Data

Production-grade RAG system over live GitHub issues with hybrid retrieval (BM25 + Qdrant + RRF), cross-encoder reranking, incremental indexing, and a RAGAS-based eval pipeline. Dockerized FastAPI service with prompt-injection protection and rate limiting.

QdrantBM25RRF RAGASFastAPIDocker
Top-5 Precision ~80% → 100%
2025
GitHub →
03
Industrial Asset Performance Analytics

SCADA time-series pipeline for wind turbine performance monitoring and power output forecasting. Classical regression combined with LSTM-based sequential modeling for real-world industrial prediction.

TensorFlowLSTM Scikit-learnSCADA
R² ≈ 0.97
2025
GitHub →
04

Skills

Languages
PythonSQLC++
Machine Learning
Scikit-learnXGBoostLightGBM Feature EngineeringPredictive ModelingClustering
Deep Learning
PyTorchTensorFlowBiLSTM CNNsTransformersNLPOpenCV
Gen AI & LLMs
LangChainLangGraphHugging Face FAISSRAGMCPPrompt Engineering
Data Science
PandasNumPyEDA A/B TestingSHAPStatistical Analysis
MLOps
FastAPIDockerMLflow ETL PipelinesGitLinuxReal-Time Inference
Platforms
AWSGCPKaggle JupyterStreamlit
05

Recognition

📄
Research Paper Under Peer Review
"VICTOR: Identity-Centric Temporal Modeling for Multi-Camera Violence Detection and Aggressor Attribution"
🏆
Smart India Hackathon (SIH)
Qualified for Internal Selection Round — national-level government hackathon
💻
200+ Coding Problems Solved
LeetCode & HackerRank — Data Structures, Algorithms, SQL
🎓
Certifications
Supervised Machine Learning — Coursera (2025)
Data Science, ML & Deep Learning — Udemy (2025)

Let's work
together.

Open to data science internships, ML engineering roles, and research collaborations. I respond within 24 hours.