Amirtha Ganesh R

Data Scientist ML Engineer GenAI & Agentic AI

I started as a traditional data scientist building ML models and MLOps pipelines, and grew into an AI engineer who builds, scales and evaluates agents, and trains LLMs from scratch with my own tokenizer.

M.Sc. in Data Science from Vellore Institute of Technology, and B.Sc. (Hons) from Sharda University, Delhi. In my previous internship at HMH TechHub, I automated HR hiring workflows with LangGraph and n8n, built LinkedIn content generation, and worked on the data preprocessing pipeline for a domain-specific LLM.

View Work Resume GitHub LinkedIn

Open to full-time roles, internships and research collaborations

Amirtha Ganesh R, data scientist
M.Sc. Data Science Vellore Institute of Technology, 2026
3 internships HMH TechHub, MBA Bazaar, White Elephant
IEEE CICN 2025 Intrusion detection paper, NIT Goa
Oracle certified OCI Data Science and Agentic AI Foundations

Skills

LLMs & AI agents
LLMs LangGraph LangChain Pydantic AI MCP RAG Corrective RAG Hybrid retrieval (BM25 + RRF) Qdrant ChromaDB n8n LangSmith BERT BPE tokenization Embeddings
Languages
Python R SQL (PostgreSQL, MySQL) MATLAB
Machine learning
Linear & logistic regression SVM XGBoost Random Forest AdaBoost K-Means DBSCAN Hierarchical clustering PCA t-SNE LDA
Deep learning
PyTorch TensorFlow Keras CNNs (ResNet, VGG, MobileNet) LSTM / GRU GANs (DCGAN, CTGAN) Transfer learning Fine-tuning
Data & visualization
Pandas NumPy Scikit-learn Matplotlib Seaborn Plotly Tableau Power BI Hypothesis testing EDA & feature engineering
MLOps & deployment
Docker FastAPI Flask Streamlit Git & GitHub MLflow Weights & Biases Oracle OCI AWS Render
Specialties
NLP & sentiment analysis Computer vision (YOLOv8) Anomaly detection Time series (ARIMA, Prophet) Hadoop Hyperparameter tuning

Projects

GoAnywhere AI travel planner
Live · Multi-agent
GoAnywhere AI

A multi-agent travel planner built on LangGraph. Flight, Hotel, Itinerary and Final Response agents share state to turn one natural-language request into flights, hotels, a day-by-day plan and a budget. An input guardrail blocks non-travel requests before any API or LLM call, and PostgreSQL checkpointing gives it persistent memory so each trip's state survives across sessions.

💡 Insight

LangGraph's PostgresSaver checkpoints every run to PostgreSQL under a thread ID, so a conversation resumes where it left off instead of starting over. The guardrail runs first, so off-topic requests never trigger paid tool or LLM calls.

LangGraph FastAPI PostgreSQL LangSmith
GitHub Live Demo
Corrective RAG with Human-in-the-Loop
Corrective RAG · HITL
Corrective RAG with Human-in-the-Loop

A privacy-aware RAG system that retrieves from private documents with hybrid search (semantic vectors + BM25, merged by Reciprocal Rank Fusion), then has an LLM evaluate whether the evidence is actually sufficient. If information is missing, it asks the user for approval before running an optional Tavily web search, and returns a grounded answer with citations and an explicit list of what's missing.

💡 Insight

Instead of guessing when retrieval falls short, the system routes on evidence quality. Private-document and public-web findings stay separated, and the web is only used with human consent.

RAG Hybrid Search BM25 + RRF Tavily
GitHub
Tamil LLM 23M
Built from scratch
Tamil LLM (23M Parameters)

A Tamil language model built end to end in PyTorch: a custom 32K BPE tokenizer, a decoder-only Transformer (6 layers, 384 hidden, 6 heads), pretraining on about 58M tokens from 100K documents, and a small SFT run. Weights and tokenizer are on Hugging Face.

💡 Insight

Experimental, and documented honestly: best pretraining validation loss was about 5.13, but generations are still repetitive. The next step is a larger corpus and 20+ epochs of pretraining.

PyTorch Transformer BPE Hugging Face
Hugging Face
Multi-Model PDF Assistant
RAG · Multi-LLM
Multi-Model PDF Assistant

Upload one or more PDFs, choose your LLM provider, and chat with your documents through a RAG pipeline. Documents are chunked and embedded locally, retrieved from a Chroma vector store, then answered by Groq or Google Gemini.

💡 Insight

Letting users switch between Groq and Gemini per session, with API keys never stored, made it a reusable tool rather than a single-provider demo.

RAG LangChain ChromaDB
GitHub Live Demo
AG's GPT agentic chatbot
AG's GPT

An agentic chatbot on a LangGraph tool-calling loop. It can search the web with Tavily, answer questions from uploaded documents using RAG on ChromaDB, remember past conversations, and stream replies through a custom black-and-white UI. Gemini is the LLM, served via FastAPI.

💡 Insight

The agent decides per message whether it needs web search, document retrieval, or no tool at all, using LangGraph's tools_condition and ToolNode instead of always calling the LLM directly.

LangGraph ChromaDB Gemini FastAPI
GitHub
Multi-Agent Research System
Multi-Agent Research System

A local multi-agent research system built with LangChain. A Search Agent and Reader Agent gather and scrape sources, a Writer Chain drafts the report, and a Critic Chain reviews it before it's finalized.

💡 Insight

A dedicated Critic Chain caught weak or unsupported claims that a single-pass LLM report would have shipped as-is.

LangChain Multi-Agent Web Scraping
GitHub
AI Automation with n8n workflows
AI Automation with n8n

Four autonomous n8n workflows: a multi-agent research war room, a job-fit and resume matcher, a company intelligence agent, and a crawler that turns web pages into structured data.

💡 Insight

Each workflow gives agents a job that ends in a decision or structured output. The resume matcher, for example, returns APPLY, CONSIDER or SKIP instead of just text.

n8n AI Agents OpenAI Tavily
GitHub
FIFA 2026 World Cup Outcome Prediction
Called the top 4 & the winner
FIFA 2026 World Cup Outcome Prediction

Two Poisson regression models predict home and away goals from historical matches, powering a simulator that runs the full 48-team tournament 1,000 times to give stage-by-stage qualification probabilities for every nation. Before the tournament started, it correctly predicted the top four teams and the eventual winner.

💡 Insight

1,000 full-tournament simulations turn match-level goal predictions into qualification odds, a more honest answer than a single bracket guess for a 48-team format.

Poisson Regression Monte Carlo Python
GitHub
Hybrid GAN Data Synthesizer
Hybrid GAN Data Synthesizer

Dual-engine synthetic tabular data generator using Gaussian Copula and CTGAN with automated preprocessing and comprehensive evaluation metrics.

💡 Insight

Synthetic data matched real distribution statistics closely enough to drop into existing pipelines without retraining downstream models.

GANs Streamlit Python
GitHub
SmartFit AI
SmartFit AI

ML-powered fitness platform analyzing 20K+ records using PCA, K-Means clustering, and neural networks for personalized insights.

💡 Insight

Clustering uncovered 4 distinct fitness personas in the data, enabling recommendations a generic dashboard would have missed.

Neural Networks K-Means Streamlit
GitHub Live Demo
RippleWorks
RippleWorks

Water quality intelligence system with dual XGBoost models for WQI regression and classification, deployed as an interactive web app.

💡 Insight

Pairing a regression score with a separate safety classifier caught borderline-unsafe samples a single model alone would have scored as fine.

XGBoost KNN Streamlit
GitHub
Sentiment Analysis RNN
Sentiment Analysis RNN

Sequential neural networks (RNN, LSTM, GRU) for text classification with real-time prediction on news articles and reviews.

💡 Insight

Gated architectures (LSTM/GRU) held onto context in longer reviews far better than vanilla RNNs, where early signal kept fading out.

LSTM GRU NLP
GitHub Live Demo
Drone Livestock Detection
Drone Livestock Detection

YOLOv8-based livestock detection system for identifying and counting sheep in aerial imagery. mAP50: 0.782, Precision: 0.793.

💡 Insight

0.782 mAP50 held up on real aerial footage, not just curated benchmark images, which is where most detection models quietly fall apart.

YOLOv8 Computer Vision UAV
GitHub
Aircraft Engine Anomaly Detection
Aircraft Engine Anomaly Detection

Identifying engines operating at abnormally high turbine gas temperatures using Isolation Forest and unsupervised machine learning.

💡 Insight

Isolation Forest flagged abnormal turbine temperatures earlier than a fixed-threshold alert would have, catching drift before it became failure.

Isolation Forest Anomaly Detection PCA
GitHub
Trader Behavior vs Market Sentiment
Trader Behavior vs Market Sentiment

Analyzing the relationship between trader profitability and the Bitcoin Fear & Greed Index using 211K+ trades. Fear periods yield 2.5× higher PnL.

💡 Insight

Fear periods produced 2.5× higher PnL than greed periods, a contrarian signal buried inside 211K+ trades most traders never look for.

Statistical Analysis Trading Analytics Behavioral Finance
GitHub
Diabetes Readmission Prediction
Diabetes Readmission Prediction

Predicting hospital readmission using ensemble learning on 101,766 patient records. Gradient Boosting achieves 69.82% accuracy.

💡 Insight

Gradient Boosting hit 69.82% accuracy across 100K+ patient records, enough to meaningfully flag high-risk readmissions before discharge.

Gradient Boosting Healthcare Classification
GitHub Dashboard
More on GitHub →

Achievements & Certifications

IEEE CICN 2025 Publication Certificate
Oracle Cloud Data Science Professional
Agentic AI Certified Foundations Associate
HP LIFE Data Science & Analytics
University of Washington Machine Learning Foundations
Research Publication

IEEE CICN 2025 Conference Publication

17th IEEE International Conference | NIT Goa | December 2025

Wrote research paper on "Dimensionality Reduction For Intrusion Detection Using Crow Search Optimization And Grey Wolf Optimization" at the prestigious IEEE International Conference on Computational Intelligence and Communication Networks (CICN-2025) held at NIT Goa, India. The research demonstrates advanced optimization techniques for enhancing cybersecurity through intelligent feature selection.

Research Publication Optimization Algorithms Intrusion Detection Dimensionality Reduction Cybersecurity
Cloud Certification

Oracle Cloud Infrastructure 2025 Certified Data Science Professional

Oracle University | September 2025

Demonstrated expertise in designing, building, and managing data science solutions on Oracle Cloud Infrastructure. Mastered OCI Data Science service, machine learning workflows, and cloud-based model deployment strategies.

OCI Cloud ML Model Deployment MLOps
Agentic AI

Agentic AI Certified Foundations Associate

Oracle | 2026

Earned Oracle's foundational certification in agentic AI systems, covering core concepts in autonomous agent architecture, orchestration, and responsible deployment — directly underpinning the production agent work built at HMH TechHub with LangGraph and Pydantic AI.

Agentic AI Agent Architecture Orchestration
Professional Development

HP LIFE: Data Science & Analytics

HP LIFE Foundation | September 2025

Completed HP LIFE's online course covering fundamentals of data science and analytics, applying core concepts of data-driven decision-making.

Data Science Fundamentals Analytics
Academic Course

Machine Learning Foundations: A Case Study Approach

University of Washington, Coursera | September 2025

Completed foundational machine learning course from University of Washington covering supervised learning, regression, classification, and case studies with real-world applications.

Machine Learning Regression Classification Python

Journey

Education & research
2024–2026

M.Sc Data Science – VIT, Andhra Pradesh

Completed M.Sc. in Data Science, specializing in statistical modeling, deep learning, and production ML/AI systems.

2025–2026

IEEE publication & certifications

Published "Dimensionality Reduction for Intrusion Detection Using Crow Search & Grey Wolf Optimization" at IEEE CICN 2025, NIT Goa. Earned Oracle Cloud Infrastructure Data Science Professional (2025) and Agentic AI Certified Foundations Associate (2026).

2020–2023

B.Sc Data Science & Analytics – Sharda University, Delhi

Graduated with Honors. Competed in 3 hackathons and founded an early-stage data-driven startup venture.

Experience
Jan–May 2026

LLM Engineer Intern – HMH TechHub, Chennai

Built an ETL pipeline turning 10+ PDFs into 410K+ tokenized units and 768-dimensional embeddings for retrieval, designed an end-to-end agentic workflow in LangGraph, delivered a RAG system for analytics reporting, and engineered type-safe workflows with Pydantic AI and Docker.

Jun–Aug 2025

Machine Learning Engineer Intern – MBA Bazaar, Bangalore

Built XGBoost and Logistic Regression models predicting trial-to-paid conversion (84% accuracy), a collaborative-filtering recommender that lifted click-through rate by 22%, and improved course completion by 18% through drop-off analysis of 15K+ records.

Jun–Jul 2022

Data Analyst Intern – White Elephant, Bangalore

Analyzed 100K+ records, improving data quality by 85% and model accuracy by 12% via PCA/t-SNE-based feature reduction.

Let's Connect

Open to collaborations, research opportunities, internships, and full-time roles in data science.

Email GitHub LinkedIn +91 6374707500