// the origin story of an ML pipeline

Every model starts messy.
So did I.

I'm Nitya — a B.Tech CS (AI/ML) student, two-time published researcher, and someone who has broken far more pipelines than she's shipped. This is the log of how the pipelines started working.

Get in touch
chapter one — ordinary world

Where the trail starts

Third-year B.Tech Computer Science (AI/ML) student at Dronacharya Group of Institutions, under AKTU — graduating 2027. Somewhere between semester exams and 2 a.m. debugging sessions, "I should learn ML" turned into two published papers and a shelf of half-finished side projects that mostly finished anyway.

Machine learning wasn't a fall — it was curiosity about how computers could learn, quietly moving in and never signing a lease to leave.

chapter two — tools gathered

What's in the pack

No hero travels empty-handed. These are the tools picked up along the way — some forged in coursework, most forged in production bugs at 1 a.m.

Python
SQL
TensorFlow / Keras
scikit-learn
BERT fine-tuning
LightGBM · XGBoost
FastAPI · Flask
Firebase
Streamlit
Pandas · NumPy
NLP
Computer Vision
DSA · LeetCode patterns
chapter three — the road of trials

Every project was a boss fight

Twelve build logs. Some shipped clean, most fought back — protobuf conflicts, MySQL errors, a PowerShell that refused to cooperate. All of them taught something the tutorials didn't.

Emotion Detection — Wellness Chatbot

★ flagship

A full-stack emotional wellness chatbot: a fine-tuned BERT classifier scores PHQ-9 signals, a Qwen2.5-based layer holds the conversation, deployed on FastAPI + Streamlit + Firebase.

BERTQwen2.5FastAPIStreamlitFirebase
→ view repo

WayzUp

civic tech

A civic-tech mapping tool built to surface and route around local issues, combining geospatial data with a simple, usable interface.

FlaskStreamlitSQLiteFoliumgeopy
→ view repo

Solar Panel Condition Classifier

computer vision

Transfer-learning model on EfficientNetB0 classifying solar panel condition across 6 classes from ~900 images — catching degradation before it costs output.

TensorFlowKerasEfficientNetB0
→ view repo

Crop Production Prediction

data science

Streamlit app forecasting crop yield using FAOSTAT data, benchmarked across Random Forest, Gradient Boosting, and Linear Regression.

Random ForestGradient BoostingStreamlit
→ view repo

Tourism Attraction Analytics

multi-task ml

A multi-task system predicting, classifying, and recommending tourism attractions in one pipeline, using LightGBM and XGBoost.

LightGBMXGBoostRecommendation
→ view repo

Insurance Claim Prediction

classification

Predictive model estimating the likelihood of insurance claims from policyholder data.

scikit-learnPandas
→ view repo

Agricultural Anomaly Detection

data science

System for flagging irregular patterns in agricultural data — built to catch the outliers that skew every downstream decision.

Anomaly DetectionPython
→ view repo

Tennis Analysis

computer vision

Video-based analysis project tracking player and ball movement to pull structured stats out of raw match footage.

OpenCVPython
→ view repo

Bird Observation

data analysis

Exploratory analysis of bird observation data — patterns, species distribution, and trends over time.

PandasVisualization
→ view repo

Food App

full-stack

A food ordering / discovery application built end-to-end as a full-stack exercise.

Full-stack
→ view repo
chapter four — allies & artifacts

The proof left behind

Two papers, two rooms full of reviewers, two "yes, publish this."

View ICOTET 2024 paper
ICOTET 2024

Customer Segmentation

Conference paper on customer segmentation techniques — presented at ICOTET 2024.

View paper
IC-AIDA 2026 · IEEE co-sponsored

Blockchain–AI for Cybersecurity

Random Forest–based anomaly detection across 14,133 scenarios, hitting 95.68% accuracy — combining blockchain integrity guarantees with AI-driven threat detection.

View paper