Coding Problem Packs

Implement-from-scratch and debugging problems for ML, computer vision, RL, transformers, generative models, and 3D — each pack is a single self-contained page with worked solutions.

9 packs 613 problems ★ ratings mark interview frequency within each pack's specialty

Comprehensive packs3

The three broad sets — start here.

Machine Learning

Numerics, activations, layers, normalization, optimizers, attention, losses, and modern tricks — the generalist ML implementation set.

93 problems14 sections★ 45 high-frequency
Sections (14)
  • Numerical foundations
  • Activations
  • Layers
  • Normalization
  • Optimizers and schedulers
  • Attention and Transformers
  • Tokenization, sampling, decoding
  • Embedding similarity and retrieval
  • Vision basics
  • Loss functions and divergences
  • Diffusion and generative
  • Reinforcement learning
  • Modern tricks: PEFT, quant, MoE
  • Data, training, miscellaneous
Open pack →Markdown source

Computer Vision

Classical vision through ViTs: filtering, geometry and warping, RANSAC, detection, segmentation, optical flow, tracking, and metrics.

92 problems16 sections★ 16 high-frequency
Sections (16)
  • Image fundamentals
  • Filtering and edges
  • Geometry and warping
  • Feature detection and matching
  • RANSAC and robust estimation
  • Object detection utilities
  • Segmentation
  • Optical flow and stereo
  • Tracking
  • Multi-view and 3D
  • Vision Transformers
  • Metrics and evaluation
  • Augmentation
  • Saliency and explainability
  • NeRF / 3DGS / volume rendering helpers
  • Practical pipelines
Open pack →Markdown source

Reinforcement Learning

MDPs and bandits through policy gradients, continuous control, offline RL, and RLHF — tabular methods to LLM-specific RL.

82 problems15 sections★ 25 high-frequency
Sections (15)
  • MDP foundations
  • Multi-armed bandits
  • GridWorld and tabular methods
  • TD with eligibility traces
  • Function approximation
  • Replay and deep value methods
  • Policy gradient
  • Continuous control
  • Model-based and planning
  • Exploration
  • Imitation learning
  • Offline RL
  • Multi-agent and planning
  • RLHF and LLM-specific RL
  • Engineering and utilities
Open pack →Markdown source

Specialized packs5

Deep dives per domain.

Transformers, Long Context & MoE

Tokenizers, attention variants, long-context positional encodings, MoE routing, decoding strategies, and state-space alternatives.

60 problems10 sections★ 32 high-frequency
Sections (10)
  • Tokenization
  • Core attention and Transformer blocks
  • Attention variants
  • Long-context positional encodings
  • Activations and FFN variants
  • Autoregressive generation and decoding
  • Mixture-of-Experts
  • Training tricks
  • State-space and alternatives
  • Closing tips
Open pack →Markdown source

Diffusion & Flow Matching

Noise schedules, DDPM and DDIM, score matching, classifier-free guidance, ODE solvers, flow matching, rectified flow, and consistency.

69 problems16 sections★ 23 high-frequency
Sections (16)
  • Noise schedules and SDE coefficients
  • DDPM forward and posteriors
  • DDPM training
  • DDPM sampling (reverse process)
  • DDIM and accelerated samplers
  • Score-based formulation
  • Classifier-free guidance and conditioning
  • Higher-order ODE solvers
  • Latent diffusion
  • Flow matching
  • Rectified flow
  • Normalizing flows
  • Mean flow and consistency
  • Architecture pieces and training tricks
  • Inversion, editing, and image-to-image
  • Closing tips
Open pack →Markdown source

KV Cache, Quantization, PEFT & Deployment

Inference efficiency end to end: KV cache, quantization, distillation, LoRA and PEFT, pruning, speculative decoding, batching, and serving.

55 problems13 sections★ 18 high-frequency
Sections (13)
  • KV cache fundamentals
  • KV-cache quantization
  • Knowledge distillation
  • Quantization fundamentals
  • Activation-aware quantization
  • PEFT: LoRA family
  • Other PEFT methods
  • Pruning
  • Speculative decoding
  • Batching and serving
  • Export and runtime
  • Distributed inference
  • Closing tips
Open pack →Markdown source

Neural Rendering, NeRF & 3DGS

Cameras and rays, volume rendering, NeRF training, Gaussian splatting, spherical harmonics, 4D/dynamic scenes, and SLAM.

60 problems14 sections★ 15 high-frequency
Sections (14)
  • Camera, rays, and geometry
  • Volume rendering
  • NeRF building blocks
  • NeRF training
  • Tri-plane / TensoRF
  • Gaussian Splatting fundamentals
  • Spherical harmonics
  • 3DGS training mechanics
  • 2DGS, surfaces, and meshes
  • Dynamic / 4D Gaussian Splatting
  • Compression and serving
  • Metrics and tools
  • SLAM and 3DGS
  • Closing tips
Open pack →Markdown source

Video, VLA & World Models

Video tokenizers and 3D VAEs, video diffusion, vision-language-action policies, and world models.

40 problems6 sections★ 15 high-frequency
Sections (6)
  • Video tokenization and VAEs
  • Video diffusion
  • VLA: Vision-Language-Action models
  • World models
  • Multimodal generation tricks
  • Closing tips
Open pack →Markdown source

Debugging drills1

Find-the-bug problems rather than implement-from-scratch.

ML Debugging Interview Problems

Realistic “this training run is broken — find it” problems across PyTorch plumbing, transformers, generative stacks, post-training, and serving.

62 problems6 sections
Sections (6)
  • Classic ML & PyTorch fundamentals (Problems 1–12)
  • Modern generative stacks: Transformers, Pre-training, Diffusion, Flow Matching, Post-training (Problems 13–22)
  • More classic ML & PyTorch (Problems 23–34)
  • More transformers & pre-training (Problems 35–42, 51–54)
  • Generative models: deeper cuts (Problems 43–46, 55–58)
  • Post-training, evaluation & serving (Problems 47–50, 59–62)
Open pack →Markdown source