Clinical Readmission Risk Modeling
A leakage-safe 30-day readmission pipeline with patient-disjoint cohorts, calibration, missingness analysis, and drift checks.
Held-out test: 0.661 ROC-AUC, 0.175 PR-AUC, and 0.077 Brier score on 10,822 encounters.Yogesh Kuchimanchi · Data Scientist
I got into data by building PCs.
I’m Yogesh, a data science graduate student at RIT. Collecting parts got me interested in the numbers behind a decision. These days, that curiosity takes me into machine learning, language, and how things change over time.
A little about me
It started with collecting parts.
I like building PCs. Somewhere between collecting parts and comparing them, I started enjoying the comparisons as much as the build itself. Why this component? What am I getting for the difference in price? Will it make a difference for what I want to do?
That is how I got interested in data. I’m now studying data science at Rochester Institute of Technology, and those questions have followed me into my projects.
What am I actually comparing?
Choosing parts means making trade-offs. A bigger number on a spec sheet only tells you so much; the workload and the rest of the build matter too. That is a question I now bring to models: did the comparison give each one a fair test?
Which information is useful?
Collecting parts also means collecting information: specs, reviews, recommendations. Making sense of all that is part of what drew me to data. With text, the same problem gets much bigger.
Does the answer change over time?
A PC build is a choice made at a particular moment. Workloads change, and what was enough before may not be enough later. I’m interested in that with models too: what happens when the data they meet changes?
And when the data is about people?
At RIT, my research takes that interest in change into a very different setting: public conversations about women’s safety in India. The question is how those conversations change after an incident, and what remains months or years later.
One question, at different moments
Time is part of the data.
These are the three time windows in our women’s safety research. Choose a window to see the question it helps us ask.
The crossroads
I want to keep doing both.
Research gives me room to investigate a question properly. Engineering lets me turn what I learn into something someone can use. As I finish my master’s, I’m looking for a data science role where I can keep asking questions and building things.
Featured projects
Here’s what those questions became.
Four projects to explore. There’s a small PC build below if you want to try it; each project is also linked directly underneath.
Explore all GitHub projectsA nod to where it started
Put the parts together.
A small interactive build, with each part representing a step in an ML project.
0 / 4 connected
DATA
FLOW
FLOW
POWER
Assembly mode
Select a component, then fit it into the matching slot.
Installed applications
Complete the build to unlock the demos.
01
Application lockedInstall all four components
×02
Application lockedInstall all four components
×03
Application lockedInstall all four components
×04
Application lockedInstall all four components
×NewsSnap: Transformer News Classification System
A four-class AG News system with a DistilBERT training pipeline, FastAPI service, React dashboard, and explicit inference modes.
Held-out evaluation: 0.870 accuracy, 0.869 macro-F1, and 0.827 MCC on 12,000 articles.Model Behavior Under Distribution Shift
A deterministic simulator for performance decay, class-prior shift, PSI, approximate KS, and configurable monitoring alerts.
Verified mixed-shift run: ROC-AUC fell from 0.758 to 0.372; max PSI reached 0.608; three alerts fired.Women-Safety Public Discourse Research
An interactive research dashboard covering 351,501 Reddit and YouTube comments across 16 women-safety cases in India.
Paper accepted at ASONAM 2026. The dashboard separates submitted-paper findings from later model audits.The full collection
All GitHub Projects.
Explore my public repositories, from machine learning systems to experiments and tools. New public repositories appear automatically.
46 of 46 repositories · Most recently updated first · Refreshing from GitHub…
rag evaluation deployment api
Deployable FastAPI RAG evaluation API with source citations, vector retrieval, Docker, and reproducible metrics.
ai engineering lab
Inference cost planning, PyTorch classification, document retrieval, and LLM evaluation tools.
NewsSnap
Full-stack news aggregator with DistilBERT summarization, React dashboard, CI/CD pipeline, and Docker deployment
readmission risk audit
Explore the source code and project files on GitHub.
nyc 311 service monitor
Explore the source code and project files on GitHub.
product analytics experimentation lab
Product analytics and A/B testing project with Python, SQL, Streamlit, and reproducible metrics.
driftlab model monitoring simulator
Explore the source code and project files on GitHub.
self healing pipeline
Bounded self-healing data pipeline with LangGraph, strict contracts, isolated repair, durable recovery, and an evidence-based technical report.
womens safety discourse dashboard
Explore the source code and project files on GitHub.
Zinga18018
Explore the source code and project files on GitHub.
portfolio website
A modern, interactive portfolio website built with Next.js, TypeScript, and Tailwind CSS
DataGuard AI
FixMyData is a smart AI-powered application designed to automatically detect and explain data quality issues in CSV files including missing values, outliers, duplicate entries, type mismatches, and more.
ML Portfolio Projects
Explore the source code and project files on GitHub.
rag document brain
Retrieval-augmented generation pipeline with sentence-transformer embeddings, ChromaDB vector search, and TinyLlama synthesis
ai code reviewer
FastAPI code review server powered by TinyLlama-1.1B with streaming SSE, severity classification, and KV-cache optimization
image captioner
Image-to-text generation using ViT encoder and GPT-2 decoder with beam search, served via FastAPI
sentiment engine
Real-time sentiment analysis API with DistilBERT, batch processing, and trend visualization via FastAPI and Streamlit
multi agent orchestrator
Multi-agent LLM framework with planner, domain experts, and synthesizer using chain-of-thought reasoning
credit card default prediction
A machine learning pipeline to predict credit card default, exposed as a RESTful API using FastAPI.
restaurant analytics llm
AI-powered restaurant analytics dashboard with natural language SQL queries. Built with Streamlit, SQLite, and LLM integration for intuitive data exploration and insights.
ML Model Monitoring and Data Drift Detection
Real-time model monitoring dashboard with KS test, PSI, and Jensen-Shannon divergence for data drift detection
Neural Style Transfer and Generative Art Lab
Neural style transfer with VGG19 and procedural generative art including fractals, flow fields, and wave interference
Knowledge Graph Construction and Reasoning Engine
NER-driven knowledge graph construction with interactive PyVis visualization and community detection
Time Series Forecasting with Transformer Architectures
Multi-model time series benchmarking with ARIMA, Holt-Winters, and automated feature engineering
Multimodal Anomaly Detection with Synthetic Augmentation
Ensemble anomaly detection with Isolation Forest, LOF, DBSCAN, and synthetic data augmentation for imbalanced datasets
Ipynb Project Snippets
Explore the source code and project files on GitHub.
SymptomAid AI
Professional AI-powered symptom analysis tool using Ollama and Streamlit. Provides structured medical insights with comprehensive safety disclaimers for educational purposes.
stock sentiment dashboard
Real-time dashboard to track public sentiment around stock tickers and compare it with actual price trends using NLP and Yahoo Finance data.
MetalVision AI
Advanced Metal Surface Defect Detection System using PyTorch with CNN, Attention Mechanisms, Ensemble Learning, SMOTE, and Cross-Validation
Hecs Refactoring
Explore the source code and project files on GitHub.
Job Hustles
Explore the source code and project files on GitHub.
imdb sentiment analysis deep learning
🎬 Advanced Deep Learning for IMDB Movie Review Sentiment Analysis - Achieving 90%+ accuracy with LSTM, BiLSTM, CNN, and Hybrid models
Wispr Clone
Explore the source code and project files on GitHub.
NeuroCanvas
🎨 An AI-powered creative platform that generates art, analyzes emotions, and creates narratives through an intuitive web interface. Built with React, Flask, and multiple AI APIs.
Project Aegis
Explore the source code and project files on GitHub.
NeuroCanvas AI Art Generator
🧠🎨 AI-Powered Emotional Art Generation Platform - Transform emotions into stunning neuromorphic art through advanced multimodal AI analysis with contextual memory and narrative generation
time chronicles ai
An immersive AI-powered historical storytelling web application that transports users through different eras with multimodal narratives, user authentication, and interactive features.
ai slide deck generator
AI-Powered Slide Deck Generator - Create professional presentations from text using Google Gemini AI with multiple styles, charts, and speaker notes
echo muse therapeutic storytelling
🎭 Echo-Muse: AI-Powered Therapeutic Storytelling Companion - Personalized healing through AI-generated stories and ambient soundscapes
sonic symbiosis
AI-powered platform that generates therapeutic soundscapes from plant bioacoustic data
mycelial memories
An AI-powered exploration of emotional connections through historical letters, visualized as living fungal networks
Python Script
Python Script to check file duplicacies
ollama image recognition tool
A Flask web application for image recognition using Ollama and a multimodal model.
ml iris data analysis
Machine Learning project for ml-iris-data-analysis
ml digits exploration
Machine Learning project for ml-digits-exploration
customer churn prediction
Chrun Based Prediction
Experience
Work that made the models more defensible.
Graduate Researcher, Data Science
Rochester Institute of Technology
- Built Python and Pandas pipelines to analyze 351,501 Reddit and YouTube comments across 16 women-safety cases in India.
- Applied Mann–Whitney U, chi-square, and G-tests; fine-tuned Qwen3.5-9B with LoRA and evaluated 3,000 held-out examples.
- Authored a paper accepted at ASONAM 2026 and documented model limitations through high-confidence error review.
GCCIS Technical Assistant
Rochester Institute of Technology
- Support 200+ Windows and Linux lab systems, troubleshoot access and software issues, and maintain Excel/VBA reports.
Research Assistant, Data Science
Rochester Institute of Technology
- Refactored Pandas and NumPy scripts and validated PostgreSQL and Excel data for accurate weekly research reporting.
Education
Built layer by layer.
M.S. in Data Science
Rochester Institute of Technology · Rochester, NY
Machine Learning, Deep Learning, Cloud Computing, Big Data AnalyticsPostgraduate Program in Data Science
Vellore Institute of Technology · India
B.C.A. in Computer Applications
New Shores International College · India
Skills
The tools on my bench.
Grouped by the work they support, not by logo count.
Languages
Python · SQL · R · Java
Machine Learning & NLP
PyTorch · scikit-learn · CatBoost · XGBoost · Transformers · LoRA · DistilBERT · FT-Transformer
Statistics & Evaluation
A/B testing · Mann–Whitney U · Chi-square · G-test · ROC-AUC · PR-AUC · Calibration · PSI / KS / JS
Data & Databases
Pandas · NumPy · SciPy · Spark · PostgreSQL · SQLite · Window functions
ML Engineering
FastAPI · Streamlit · React · Docker · GitHub Actions · Automated testing · Git · Excel / VBA
Contact
Have a real data problem?
I’m open to data science and ML engineering opportunities, research collaboration, and useful technical conversations.