Hi, I'm

Kiran Babu Athina

Data Scientist & Machine Learning Engineer

I build

From knowledge-graph RAG on local GPUs to PySpark pipelines moving 100M+ records a day, I turn complex data into models that hold up in production, and every system comes with an eval set and a baseline.

Portrait of Kiran Babu Athina

Rigorous evaluation, production-grade AI

I'm a Data Scientist and Machine Learning Engineer with four years of experience turning large, complex datasets into statistical models, predictive systems, and applied ML/GenAI solutions across telecom, finance, healthcare, agriculture, and scientific research.

At Texas A&M's AISLS Lab, I build knowledge-graph RAG systems that answer questions from research papers and cite the exact source page, running entirely on local GPUs. Before that, at Tata Consultancy Services, I built PySpark pipelines, NLP classifiers, and predictive models deployed with Docker and Kubernetes.

I hold an M.S. in Data Science from Texas A&M (3.9 GPA). I don't call a model done until it has a held-out eval set, a baseline to beat, and a clear recommendation stakeholders can act on.

4+
Years building production ML
95%
RAG faithfulness on a 10K-question eval
100M+
Records/day through PySpark ETL
3.9
GPA, M.S. Data Science

Where I've worked

Research Assistant

Texas A&M AISLS Lab · College Station, TX

Jun 2025 – Present
  • Engineered a triple-retriever RAG pipeline (KG semantic, Qdrant dense, BM25) with Reciprocal Rank Fusion and cross-encoder reranking, achieving 95% faithfulness and 89% answer accuracy on a 10,000-question domain eval set.
  • Boosted context recall from 0.38 → 0.94 by constructing a domain knowledge graph of 7,405 entities and 11,249 relationships from 235 research papers, using a multi-pass Qwen3 14B extraction pipeline with 93% verbatim quote verification.
  • Developed an agentic retriever with 8 targeted graph and passage search actions over a Kuzu graph database and embedded Qdrant index, achieving 91% gold-page retrieval and 86% gold-quote recall, measured deterministically against source PDFs.
  • Designed a React frontend with clickable [F#]/[P#] citations that open the exact source PDF page with the retrieved passage highlighted and per-citation retrieval scores shown, for full end-to-end citation transparency.
Knowledge GraphsAgentic RAGKuzuQdrant BM25 + RRFCross-encoderQwen3React

Graduate Research Assistant

Texas A&M Department of Animal Science · College Station, TX

Jan 2024 – May 2025
  • Enabled minute-level per-animal energy tracking and monthly methane forecasting, validated against live MQTT sensor ground truth, by engineering a sensor-driven digital twin with a Python Mesa agent-based model on university HPC infrastructure.
  • Eliminated dependency on general-purpose cloud APIs for livestock domain queries, serving 3 research semesters of faculty-validated production use, by fine-tuning LLaMA 3 with LoRA/PEFT on a proprietary animal science corpus and integrating Azure AI Search with Azure OpenAI for grounded, citation-backed generation.
LLaMA 3PEFTAzure OpenAI Mesa ABMMQTT

Machine Learning Engineer

Tata Consultancy Services · Hyderabad, India

Aug 2021 – Jul 2023
  • Processed 100M+ sensor records per day for downstream ML fault prediction by designing a PySpark ETL framework on Hadoop to extract, filter, and aggregate telecom gateway logs.
  • Cut loan processing time by 30% by developing classification models on Azure ML with AutoML-driven selection and statistical feature engineering on approval probability scores.
  • Eliminated 30% of manual hardware log triage and improved issue resolution time by 18% with a TF-IDF + Random Forest NLP classification pipeline across 500,000+ daily telecom log entries.
  • Reduced post-deployment model decay by 22% and cut deployment turnaround by 30% by establishing drift-monitoring dashboards and containerizing ML applications with Docker on Kubernetes.
PySparkHadoopAzure ML NLPRandom ForestDockerKubernetes

Machine Learning Engineer Intern

Tata Consultancy Services · Hyderabad, India

May 2021 – Jul 2021
  • Achieved 98.5% fraud classification accuracy on KYC anomaly detection with a multimodal PyTorch model fusing 1D CNN audio and 2D CNN document-image embeddings.
  • Reduced onboarding and handoff time by 25% by writing end-to-end documentation of model architectures, assumptions, and evaluation results.
PyTorchCNNsMultimodalFraud Detection

My toolkit

Languages

PythonSQLJavaScriptRC

ML / AI

LLMsRAGAgentic AIKnowledge GraphsLoRAQLoRAPEFTTransformersVision-Language ModelsCLIPPrompt EngineeringHybrid Retrieval (RRF)Cross-encoder RerankingStatistical ModelingExperimental DesignTime Series (ARIMA / SARIMA / LSTM)

Frameworks

PyTorchTensorFlowHugging FaceScikit-learnLangChainLangGraphFAISSNLTKXGBoostLightGBMFastAPIReactStreamlit

Cloud & HPC

Azure AI SearchAzure AI FoundryAzure OpenAIAzure MLAzure App ServiceAzure Static Web AppsAzure Blob StorageAWS S3AWS EC2AWS SageMakerNVIDIA GPU Clusters

Tools & Data

DockerKubernetesCI/CDGit / GitHubPySparkHadoopTableauPostgreSQLMySQLMongoDBRedisQdrantKuzuOllamaMQTT

Things I've built

Research
95%faithfulness · 89% answer accuracy on a 10,000-question eval

KG-RAG: Chat with Animal Nutrition Research

Ask a question about livestock feeding and get an answer drawn from 235 research papers. It searches a knowledge graph, the papers' passages, and keywords at once, ranks what it finds, and cites the exact PDF page behind every claim.

KuzuQdrantBM25 + RRFGemma 4React
Healthcare AI
91%diagnostic accuracy across 12 retinal diseases

Retinal Disease Diagnosis with Vision-Language Models

Fine-tuned MedGemma 4B on paired fundus images and clinical notes, validated on 1,200+ expert-annotated images.

MedGemma 4BVLMsFine-tuningMedical Imaging
Deep Learning
+17%image-text retrieval on MSCOCO (+14% zero-shot on ImageNet)

Image-Text Retrieval with Contrastive Learning (CLIP-style)

Dual-encoder ResNet-50 + DistilBERT model trained with NT-Xent and iSogCLR losses using AdamW/RAdam optimizers, benchmarked against baseline CLIP.

PyTorchResNet-50DistilBERTContrastive Learning
NASA Space Apps 2025
96.64%classification accuracy on Kepler + TESS data

Exoplanet Discovery

Ensemble of Random Forest, XGBoost, LightGBM, Gradient Boosting, and MLP models that classifies Kepler and TESS objects as confirmed exoplanets, candidates, or false positives, deployed as an interactive Streamlit app.

XGBoostLightGBMScikit-learnStreamlit
Research
Minute-levelper-animal energy tracking + monthly CH₄ forecasts

Livestock Digital Twin

Sensor-driven agent-based simulation that ingests live MQTT sensor streams to model per-animal energy and methane dynamics.

PythonMesa ABMMQTTSimulation

Research output

Indexed on Google Scholar · Texas A&M University · Artificial Intelligence Scholar profile
2025
An agent-based framework of cattle value discovery system for precision nutrient requirements and utilization prediction of beef cattle

K. Kaniyamattam, V. Kulangara-Veettil, P. Kundu, K. B. Athina, L. O. Tedeschi

CABI Digital Library · Vol. 16, Issue 3, p. 449 · September 2025

Translates the deterministic Cattle Value Discovery System (CVDS) into an agent-based model, simulating individual cattle as autonomous agents with their own body weight, body condition score, and metabolic efficiency to predict daily gain, days to finish, and carcass composition.

0 Citations

Academic background

Texas A&M University

M.S. in Data Science

Aug 2023 – May 2025 · College Station, TX

GPA 3.9 / 4.0

Gayatri Vidya Parishad College of Engineering

B.Tech. in Electronics & Communication Engineering

Aug 2017 – Jul 2021

First Class with Distinction

Certifications

NVIDIA DLI Certified Generative AI Fundamentals PCAP: Programming Essentials in Python CCNA: Introduction to Networks SailPoint Identity Security Leader Intermediate SQL Queries

Let's build something

Based in Austin, TX. Send a note below and I'll get back to you.