Hello, I'm

Bhavesh Shinde

And I build machine learning systems that can explain their own reasoning.

MSc Artificial Intelligence student based in London, originally from Mumbai. I like problems where you have to be honest about what your data can and can't tell you — not just chase a leaderboard number.

My work spans NLP, explainable AI, and applied statistics: building full pipelines end to end, from sourcing and balancing imbalanced data through to validating results with bootstrap confidence intervals rather than a single point estimate.

See my work Get in touch

Skills

What I actually reach for, grouped by where it shows up in my work.

Machine learning

Predictive modelling & classification
Regression & clustering
Ensemble learning, SVMs, CNNs
Deep learning

NLP & explainability

Linguistic feature engineering
Transformer embeddings (BERT)
Text classification
SHAP-based explainable AI

Data science

Data cleaning & preprocessing
Exploratory data analysis
Bootstrap statistical validation
Data visualisation

Libraries & frameworks

NumPy, Pandas, Matplotlib
scikit-learn, XGBoost
TensorFlow, Keras, PyTorch
Hugging Face, spaCy, NLTK

Languages

Python, R
Java, SQL

Tools

Git & GitHub
VS Code
Jupyter Notebook

Notable projects

Two projects that show how I actually work: full pipelines, honest evaluation, and a result you can interrogate.

Four-Class Human–AI Collaborative Writing Detector

MSc Dissertation · 2026

Most AI-text detectors treat authorship as binary — human or AI. That framing breaks down for how students actually use ChatGPT: drafting, editing, polishing. I built a four-class classifier that separates pure human and pure AI writing from two collaborative in-between cases, and used SHAP to turn the model's output into plain-English explanations an educator could act on.

0.78 macro-recall with XGBoost on BERT embeddings, a 62% relative improvement over a published RoBERTa baseline (0.48)
Engineered 13 handcrafted linguistic features across lexical, syntactic, sentiment, and structural dimensions
Validated results with bootstrap confidence intervals and an internal binary baseline, not a single headline number

0.78

macro-recall vs. a 0.48 published baseline

Python XGBoost BERT SHAP spaCy / NLTK

Read the write-up View on GitHub

Advanced Crypto Market Intelligence System

Academic Project · ACMIS

A multi-agent pipeline that automates crypto market analysis end to end — acquiring data, cleaning it, scoring sentiment, forecasting price direction, and writing up what it found in plain language.

Combined statistical volatility metrics with transformer-based sentiment analysis into a composite risk score
Applied linear regression to forecast 5-day price trends
Integrated a pre-trained BART model to auto-generate human-readable market reports

stages automated: acquire, clean, score, report

Python Hugging Face BART Regression

View on GitHub

1 / 2