// aspiring ml engineer & builder

Utkarsh
Mishra

Turning raw data into intelligent systems. I design and train machine learning models that solve real-world problems — from vision to language to structured predictions.

View My Models → Get in Touch
12+
Projects Built
5+
Domains Explored
3+
Years Learning

01 / About

The mind behind
the models

I'm an aspiring ML engineer with a deep passion for building systems that learn, adapt, and predict. My work spans computer vision, NLP, and classical ML — always with a focus on clean experiments and reproducible results.

I believe great ML isn't just about accuracy metrics. It's about understanding the data deeply, designing thoughtful architectures, and communicating results clearly.

Currently exploring the intersection of large language models and practical applications, while sharpening my fundamentals in mathematics and systems thinking.

PyTorch
TensorFlow
Scikit-learn
Python
Transformers
Pandas / NumPy
Computer Vision
NLP
MLflow
Jupyter

02 / Models

What I've built

◈ Computer Vision
Plant Disease Detector
CNNResNet-50PyTorchTransfer Learning

Fine-tuned ResNet-50 to classify 38 plant disease categories from leaf images. Trained on the PlantVillage dataset with heavy augmentation to handle field-condition noise.

96.4%
Accuracy
38
Classes
54K
Images
◈ NLP
Sentiment Analyser
BERTHugging FaceFine-tuning

Fine-tuned BERT-base on multi-domain reviews (movies, products, restaurants). Handles nuanced sentiment including sarcasm and negation patterns better than bag-of-words baselines.

93.1%
F1 Score
3-class
Output
110M
Params
◈ Classical ML
House Price Predictor
XGBoostFeature Eng.Sklearn

End-to-end regression pipeline using XGBoost with extensive feature engineering on the Ames Housing dataset. Includes SHAP explanations for interpretability.

0.12
RMSE (log)
89
Features
Top 8%
Kaggle
◈ Computer Vision
Face Mask Detector
MobileNetV2OpenCVReal-time

Lightweight real-time mask detection using MobileNetV2 backbone. Optimised for edge inference — runs at 30fps on CPU. Includes face localisation pre-processing.

98.7%
Accuracy
30fps
Inference
4.2MB
Model Size
◈ NLP
News Topic Classifier
DistilBERTZero-shotFastAPI

Multi-label news categorisation using DistilBERT with zero-shot capability for unseen topics. Deployed as a REST API; processes 500+ articles per second.

91.5%
Precision
8
Topics
500/s
Throughput
◈ Classical ML
Customer Churn Predictor
Random ForestSMOTEExplainability

Churn prediction pipeline for a telecom dataset. Handles severe class imbalance with SMOTE + calibrated probabilities. SHAP waterfall plots explain individual predictions.

87.3%
ROC-AUC
82%
Recall
SHAP
Explain

03 / Experience

The journey so far

Building skills one experiment at a time.

2024 — Present
ML Research Enthusiast
Self-directed / Open Source

Independently replicating landmark papers (Attention Is All You Need, ResNets, DDPM). Contributing to open-source ML repositories and participating in Kaggle competitions to stress-test fundamentals.

2023 — 2024
ML Projects & Coursework
Academic / Personal Portfolio

Completed Andrew Ng's Deep Learning Specialisation and fast.ai. Built 6+ end-to-end projects covering vision, NLP, and tabular ML. Focused on clean code, reproducibility, and experiment tracking with MLflow.

2022 — 2023
Python & Data Science Foundations
Self-taught

Mastered core Python, NumPy, Pandas, and Matplotlib. Built classical ML models from scratch to understand the mathematics. First Kaggle competition: top 20% on Titanic survival prediction.

04 / Contact

Let's build something
together

Open to collaborations, internships, research opportunities, and interesting conversations about machine learning.