Rahul Ray: AI/ML · Data Science

About

Pushing methods past their assumptions.

I'm a computer science graduate and current MSc student at Leiden University, based in Den Haag. My work spans classical ML, deep learning research, and edge AI deployment, with a particular interest in the questions that sit at the boundary of what a method can actually do.

I care about methodology: controlled baselines, proper statistical testing, and honest reporting when the results don't go the way you expected. Previously ML intern at Fox Solutions working on Vision Transformers for edge deployment.

Core

Python · PyTorch pandas · NumPy · scikit-learn Experiment Design Statistical Testing · Nested CV

Classical & Tabular ML

XGBoost · LightGBM · CatBoost SHAP · Optuna Feature Engineering · Selection Wilcoxon · Cohen's d

Deep Learning & RL

Vision Transformers · ResNet / CNNs Self-Supervised Learning Masked Representation Prediction EMA / Momentum Encoders Q-Learning · DQN Policy Gradients · PPO Actor-Critic Mixup · Label Smoothing Generalisation Theory

Text Mining & NLP

sentence-transformers TF-IDF · BM25 · c-TF-IDF Topic Modelling · NER UMAP · HDBSCAN

Information Retrieval

FAISS · Dense Retrieval CLIP · Multimodal Retrieval mAP · NDCG · Recall@k Query Expansion · Re-ranking

MLOps & Deployment

Automated ML pipelines Edge deployment · Model quantization Heuristic & graph search Hugging Face · Streamlit

Projects

Selected work

Can LLMs Make Tabular ML Better?

A controlled experiment testing whether LLM-generated semantic features improve XGBoost on structured data, using nested CV, Wilcoxon tests, and Cohen's d to evaluate the hybrid approach properly. The results went against the hypothesis, and that turned out to be the more interesting finding.

XGBoost LLaMA 3.1 sentence-transformers Optuna SHAP

Read writeup → ↗ GitHub

AutoInsight: End-to-End AutoML App

A browser-based AutoML tool that trains 7 models in parallel, generates SHAP explainability, and produces LLM-written plain-language insights, with a privacy-preserving design where raw data never leaves the browser.

Streamlit XGBoost · LightGBM · CatBoost SHAP Optuna Groq API

Read writeup → ⚡ HF Space ↗ GitHub

Quake: Learned Heuristic A* on BSP Geometry

Parsed Quake 1's raw binary map format to build navigation graphs, generated 625k ground-truth pathfinding pairs, then trained learned heuristics that cut A*'s node expansions by 42% while keeping paths near-optimal.

PyTorch XGBoost BSP v29 parsing A* search NetworkX

Read writeup → ↗ GitHub

Epoch-wise Double Descent Under Label Noise

A reproduction of Nakkiran et al. (2019)'s epoch-wise double descent on CIFAR-10 with ResNet-18 under varying label noise, with a mechanistic investigation into what's actually happening at the peak.

PyTorch ResNet-18 CIFAR-10 Label Noise Generalisation Theory

Read writeup → ↗ GitHub

Newsphere Live: Semantic News Map

A self-updating 3D map of current news where proximity means semantic similarity. Fetches 11 RSS feeds every 6 hours, embeds articles via MiniLM, reduces to 3D with UMAP, clusters with HDBSCAN, and publishes to GitHub Pages, zero backend, zero cost.

sentence-transformers UMAP HDBSCAN Three.js GitHub Actions

Read writeup → ⚡ Live ↗ GitHub

JEPA-IR: Benchmarking I-JEPA for Image Retrieval

Can a model never trained contrastively still retrieve well? Benchmarks I-JEPA (ViT-H/14) as a frozen feature extractor against CLIP and ResNet-50 on Oxford5k and Paris6k. The gap, and the near-zero query expansion gain for I-JEPA, points to a fundamental embedding space issue, not a tuning problem.

I-JEPA CLIP FAISS Oxford5k · Paris6k mAP PyTorch

Read writeup → ↗ GitHub

Text-JEPA: Representation Prediction for Language

Adapts I-JEPA's masked representation prediction objective from vision to language using a DistilBERT-based context encoder, EMA target encoder, and shallow predictor. Trained on WikiText-103. Embedding geometry improved markedly (cosine similarity 0.93 → 0.60), but downstream task utility fell; the two metrics decoupled mid-training.

Text-JEPA DistilBERT Self-Supervised Learning EMA WikiText-103 PyTorch

Read writeup → ↗ GitHub

01 / 07

Experience

Where I've
been.

Work

Mar 2025 –
Aug 2025

Machine Learning Intern

Fox Solutions Pvt. Ltd. · India

Built a computer vision pipeline using Vision Transformers, working from model development through to edge deployment. Adapted ViT architectures to hardware constraints, optimising inference latency and memory footprint without significant accuracy loss.

Vision Transformers Edge Deployment ONNX PyTorch

Education

Sep 2025 –
Aug 2027

MSc Computer Science – Data Science

Leiden University · Leiden, Netherlands

2021 –
2025

BTech – Computer Science

SRM Institute of Science and Technology · Chennai, India

RahulRay

Pushing methods past their assumptions.

Selected work

Where I'vebeen.

Let's work on something worth working on.

Rahul
Ray

Where I've
been.