Writing
Selected essays, technical reports, and practical notebooks.
I have published 70 technical reports and 73 public notebooks across LLM systems, RAG, deep learning, computer vision, generative modeling, training infrastructure, and applied ML. The archive includes tutorials, research reproductions, implementation notes, and competition notebooks.
70 technical reports 73 public notebooks 3 technical courses
Bodies of work
LLM systems, RLHF, and RAG
From preference modeling and RLHF implementations to retrieval pipelines and systematic LLM evaluation.
Computer vision and generative imaging
Detection, segmentation, interpretability, GAN evaluation, limited-data training, colorization, and image editing.
Framework practice
Practical work across PyTorch, Lightning, Keras, TensorFlow, Hugging Face, and Transformers.
Training mechanics
Optimization, augmentation, class imbalance, metric learning, and convolution mechanics.
3D vision and video
Novel-view synthesis, neural scene flow, inpainting, talking heads, video retiming, and animation.
Applied experimentation
Research reproductions and end-to-end notebooks across medical imaging, object detection, audio, tabular ML, and competitions.
Selected writing
How to Evaluate, Compare, and Optimize LLM Systems
A practical evaluation workflow using human labels, LLM judges, retrieval metrics, and configuration search.
A practical evaluation workflow using human labels, LLM judges, retrieval metrics, and configuration search.
Building Advanced Query Engine and Evaluation with LlamaIndex and W&B
Tracing, versioning, and evaluating a multi-stage retrieval pipeline.
Tracing, versioning, and evaluating a multi-stage retrieval pipeline.
Understanding Reinforcement Learning from Human Feedback
An intuitive route through reward modeling and learning from preference data.
An intuitive route through reward modeling and learning from preference data.
Towards Deep Generative Modeling with W&B
A hands-on guide to building and inspecting generative models.
A hands-on guide to building and inspecting generative models.
Interpretability with CAM and Grad-CAM
Understanding what image classifiers attend to and where their reasoning fails.
Understanding what image classifiers attend to and where their reasoning fails.
COVID-19 detection with YOLOv5
An end-to-end training notebook and my most-upvoted Kaggle contribution.
An end-to-end training notebook and my most-upvoted Kaggle contribution.
How the work evolved
Applied computer vision and technical writing
Built and documented sign-language recognition, face analysis, image inpainting, interpretability, and generative-modeling projects.
Built and documented sign-language recognition, face analysis, image inpainting, interpretability, and generative-modeling projects.
Deep learning systems and research reproduction
Implemented and benchmarked work across PyTorch, TensorFlow, GANs, self-supervised learning, 3D vision, and medical imaging.
Implemented and benchmarked work across PyTorch, TensorFlow, GANs, self-supervised learning, 3D vision, and medical imaging.
LLM training and open-source infrastructure
Worked on RLHF, Transformers, framework integrations, and reproducible training.
Worked on RLHF, Transformers, framework integrations, and reproducible training.
Production RAG, evaluation, and agents
Wrote about production RAG, LLM evaluation, observability, and agents.
Wrote about production RAG, LLM evaluation, observability, and agents.
Selected collaborations
Introduction to image inpainting with deep learning
With Sayak Paul · traditional methods, autoencoders, and partial convolutions.
With Sayak Paul · traditional methods, autoencoders, and partial convolutions.
Understanding the effectiveness of ensembles in deep learning
With Sayak Paul · loss landscapes, generalization, and why ensembles work.
With Sayak Paul · loss landscapes, generalization, and why ensembles work.
Unsupervised visual representation learning with SwAV
With Sayak Paul · reproducing self-supervised representation learning in TensorFlow.
With Sayak Paul · reproducing self-supervised representation learning in TensorFlow.
An introduction to adversarial latent autoencoders
With Sairam Sundaresan · disentangled latent representations and high-quality image generation.
With Sairam Sundaresan · disentangled latent representations and high-quality image generation.
Keras XLA benchmarks
With Soumik Rakshit and Sayak Paul · measuring XLA performance across Keras training workloads.
With Soumik Rakshit and Sayak Paul · measuring XLA performance across Keras training workloads.
Complete archive
Reports and notebooks are grouped below by topic and publication venue. You can also browse my work and talks and workshops.
Weights & Biases Reports
Technical articles and tutorials I’ve written on the W&B platform—covering deep learning, computer vision, LLMs, and practical ML engineering.
LLMs & RLHF
- Understanding Reinforcement Learning from Human Feedback (RLHF): Part 1
- An Introduction to Training LLMs Using Reinforcement Learning From Human Feedback (RLHF)
- How to Evaluate, Compare, and Optimize LLM Systems
- Building Advanced Query Engine and Evaluation with LlamaIndex and W&B
- Implementing RLHF: Learning to Summarize with trlX
- RLHF: Hyperparameter Optimization for trlX
Computer Vision & GANs
- Towards Deep Generative Modeling With Weight & Biases
- How to Evaluate GANs using Frechet Inception Distance (FID)
- Train Generative Adversarial Networks (GANs) With Limited Data
- In-Domain GAN Inversion for Real Image Editing
- An Introduction to Adversarial Latent Autoencoders
- Introduction to Image Inpainting with Deep Learning
- Interpretability in Deep Learning With Weights & Biases: CAM and Grad-CAM
- An Overview of Instance Aware Image Colorization
- Image Segmentation Using Keras and Weights & Biases
- Object Localization With Keras and Weights & Biases
- Image Classification Using Vision Transformer and KerasCV
Framework practice
PyTorch & PyTorch Lightning
- How To Use GPU with PyTorch
- How to save and load models in PyTorch
- How To Check If PyTorch Is Using The GPU
- Image Classification Using PyTorch Lightning and Weights & Biases
- Transfer Learning Using PyTorch Lightning
- Multi-GPU Training Using PyTorch Lightning
- An Introduction To The PyTorch View Function
- PyTorch Dropout for regularization - tutorial
- Implementing and Tracking the Performance of a CNN in PyTorch
Keras & TensorFlow
- Keras Dense Layer: How to Use It Correctly
- Input Keras Layer Explanation With Code Samples
- Batch Normalization in Keras - An Example
- How to Use Cosine Decay Learning Rate Scheduler in Keras
- Using GPUs With Keras: A Tutorial With Code
- Integrating Keras with Weights & Biases
- How To Install TensorFlow With GPU Support on Windows
- How to Prevent TensorFlow From Fully Allocating GPU Memory
- LSTM RNN in Keras: Examples of One-to-Many, Many-to-One & Many-to-Many
- A Guide to Multi-Label Classification on Keras
- Keras XLA Benchmarks
Hugging Face & Transformers
Training mechanics
Deep Learning Fundamentals
- What’s the Optimal Batch Size to Train a Neural Network?
- ReLU vs. Sigmoid Function in Deep Neural Networks
- How to Handle Images of Different Sizes in a Convolutional Neural Network
- Intuitive understanding of 1D, 2D, and 3D convolutions in convolutional neural networks
- What’s the Difference Between Strided Convolution and Pooling?
- Comparing Sigmoid-MSE With Softmax Cross-Entropy for Image Classification
- How One-Hot Encoding Improves Machine Learning Performance
- Understanding the Effectivity of Ensembles in Deep Learning
Data Augmentation & Training Techniques
3D vision and video
3D Vision & Novel View Synthesis
Video & Animation
Applied and exploratory work
- Self-Supervised Learning with SwAV
- How To Use Weights & Biases With MMDetection
- Generating Digital Painting Lighting Effects via RGB-space Geometry
- An Overview of DeepFaceDrawing
- Rewriting a Deep Generative Model: An Overview
- From EMNIST to Wikisplit: Sentence Composition Using W&B
- Modeling Drivable Areas for Autonomous Vehicles with Real and Synthetic Data
Medium
Kaggle Notebooks
- [Train] COVID-19 Detection using YOLOv5 (800 upvotes)
- Experiment Tracking with Weights and Biases (420 upvotes)
- [Train] Brain Tumor as Video Classification + W&B (175 upvotes)
- [Train] NFL Extra Images YOLOv5 with W&B (174 upvotes)
- Brain Tumor EDA and Interactive Viz with W&B (158 upvotes)
- Train YOLOv5 + Cross Validation + Ensemble + W&B (155 upvotes)
- TPS June [Weights and Biases] (148 upvotes)
- Submission Covid19 (132 upvotes)
- [G2Net] Understand the Data (129 upvotes)
- SD 2.0: Download Dataset from W&B Artifacts (111 upvotes)
- Visualize Bounding Boxes Interactively (111 upvotes)
- [V2] Self-Supervised Pretraining with SwAV (103 upvotes)
- EfficientNet+Mixup+K-Fold using TF and wandb (63 upvotes)
- HPA: Segmentation Mask Visualization with W&B (62 upvotes)
- Comparative Study: SETI Experiments (61 upvotes)
- Create Constant Q Transformed Dataset (51 upvotes)
- HPA: Multi-Label Classification with TF and W&B (50 upvotes)
- Easy PCA TSNE and UMAP (49 upvotes)
- [Train] Study Level Classifier + W&B (49 upvotes)
- [Train] COTS Detection using YOLOv5 and W&B (49 upvotes)
- Quick Data EDA+ Segmentation Viz using W&B (49 upvotes)
- [Train][PyTorch] FineTune ConvNeXt (44 upvotes)
- [Train] TensorFlow+EfficientNet+Kfold+W&B (42 upvotes)
- Use PyTorch Lightning With Weights and Biases (41 upvotes)
- Visualize the COTS interactively with W&B (36 upvotes)
- [MLOps Tutorial] XGBoost with W&B (36 upvotes)
- TensorFlow Inference [No-TTA] (28 upvotes)
- BirdCLEF: Quick EDA with W&B (26 upvotes)
- APTOS Blindness Detection - EDA and Keras ResNet50 (25 upvotes)
- Interactive EDA using W&B Tables (23 upvotes)
- Cassava 2019+2020 Dataset Prep+Versioning w/ W&B (23 upvotes)
- GradCAM: Implementation+Visualization in TF & W&B (22 upvotes)
- XGBoost Hyperparameter Tuning with W&B Sweeps (22 upvotes)
- Train EfficientNetv2 with TF and W&B (20 upvotes)
- [HappyWhale] EDA - IDs to Images using W&B Tables (19 upvotes)
- Ubiquant EDA (18 upvotes)
- Transformer Baseline with TF/Keras and W&B (18 upvotes)
- Get the most out of LightGBM using W&B (17 upvotes)
- Sartorius: Instance Mask Viz with W&B (16 upvotes)
- [TensorFlow] Working with TIFF (15 upvotes)
- Rainforest Create Image Dataset with W&B (14 upvotes)
- Resampled Voxel Dataset - with KaggleRecipes (W&B) (14 upvotes)
- AutoEncoders implementation (13 upvotes)
- G2Net: Quick EDA and CQT + W&B (13 upvotes)
- Style Transfer using CycleGAN with TF and W&B (13 upvotes)
- BMS: EDA and Dataset Visualization w/ W&B (12 upvotes)
- Single Label Cell Label Dataset Creation w/ W&B (12 upvotes)
- Comparing Different Modeling Methods w/ W&B (11 upvotes)
- Better Data Understanding with W&B Tables (10 upvotes)
- HPA: Inference (10 upvotes)
- Investigate Imgur5k dataset (10 upvotes)
- Images to TFRecord (10 upvotes)
- Experiment Tracking w/ Weights and Biases [Russia] (10 upvotes)
- Rainforest Train Baseline Model with W&B (9 upvotes)
- All in one MLB Quick EDA with W&B Tables (8 upvotes)
- RGB 512x512 Dataset Creation+Versioning with W&B (7 upvotes)
- Save the Dataset as W&B Artifacts (7 upvotes)
- Visualize Dataset Interactively using W&B Tables (6 upvotes)
- Dataset Versioning for Single Label Dataset (6 upvotes)
- Testing W&B Offline Flow (6 upvotes)
- Batch Normalization ft Batch Size with W&B (6 upvotes)
- Jigsaw BERT Model (5 upvotes)
- Rainforest: Which model is better? (5 upvotes)
- Adaptive Gradient Clipping: Ablation Study w/ W&B (5 upvotes)
- Using W&B DSViz for Cell Level Dataset Creation (5 upvotes)
- Kaggle Survey 2020: W&B vs All (5 upvotes)
- Kaggle Survey 2021: W&B vs All (5 upvotes)
- Cell Level Extra Data Download (5 upvotes)
- Test wandbcode (4 upvotes)
- Multiprocessing 5 Second Audio Clips (4 upvotes)
- HuBMAP: TF with TPU EfficientUNetB7 512x512 [Train] (4 upvotes)
- CommonLit Readability Prize: EDA + Baseline (3 upvotes)
- Visualize the Dataset using W&B Tables (3 upvotes)