NudgeCode: an AI coding interview tutor that gives hints, not answers

NudgeCode is a practice site for software engineers preparing for coding interviews: 355 algorithm problems in 13 patterns, 29 ML coding tasks, AI infra lessons and AI mock interviews. Python and JavaScript run in your browser, and the AI tutor reads your code and the failing test, then gives a hint instead of the solution. Free to start: no credit card, no install. Pro is $19.99 a month or $129 a year.

How it works

1. Name the pattern. Guess which of the 13 patterns fits before you write a line. A wrong guess shows you the signals you missed.
2. Take a nudge. The first nudge is a question. Then the approach, then pseudocode with TODOs. Never the full solution.
3. Ask the tutor. The tutor reads your code and the failing test, then answers with a trace instead of a fix.

Algorithm patterns

• Array & Hash Table: Master array manipulation and hash table techniques for efficient lookups and storage
• Two Pointers: Learn efficient two-pointer techniques for array and string problems
• Sliding Window: Master sliding window technique for substring and subarray problems
• Binary Search: Learn binary search and its variations for efficient searching
• Linked List: Master linked list manipulation and traversal techniques
• Stack: Learn stack data structure for LIFO operations
• Binary Tree: Master binary tree traversal and manipulation
• Dynamic Programming: Master dynamic programming for optimization problems
• Graph: Learn graph algorithms and traversals
• Greedy: Master greedy algorithms for optimization
• Backtracking: Learn backtracking for combinatorial problems
• Heap: Master heap/priority queue data structure
• Bit Manipulation: Learn bit manipulation techniques

AI infra track

• Math Foundations: Matmul shapes, block matrices, softmax as a distribution, quantization error, lossless speculative decoding and the chain rule, checked with small tensors.
• Memory & FLOPs Accounting: Count parameters, FLOPs, KV cache bytes and training memory before you optimize anything.
• Numerics & Hand-Written Backward: Stable softmax, LSE, cross-entropy gradients and LayerNorm/RMSNorm backward, checked against autograd.
• Transformer Inference Building Blocks: RoPE in both layouts, GQA without copies, KV cache and sampling, checked against a full forward.
• GPU Hardware and Architecture: SMs, Tensor Cores, the memory hierarchy, peak FLOPS and the roofline, architecture generations from V100 to B200, and NVLink vs PCIe.
• GPU Programming Model, Simulated: Threads, warps, shuffles, coalescing, bank conflicts, occupancy and tiled GEMM, simulated in PyTorch.
• Online Softmax & FlashAttention: Merge partial softmax states, then build FlashAttention v1, v2 and Flash-Decoding block by block.
• Paged KV Cache: Slot mapping, block tables, paged and varlen attention, and a nano-vllm style BlockManager.
• Scheduling & Prefix Caching: Continuous batching, preemption, chunked prefill and radix-tree prefix caching.
• Speculative Decoding: Draft, verify and accept tokens without changing the target distribution.
• Quantization: INT8/INT4/FP8 schemes, SmoothQuant, GPTQ-style error compensation and KV cache quantization.
• Distributed Training, Simulated: Collectives, ring all-reduce, DDP, ZeRO, tensor and pipeline parallelism on a list of fake ranks.

More

• All 355 problems
• ML coding: Write attention, KV cache, schedulers and loss functions from scratch, checked by tests in your browser.
• AI mock interview
• Spaced review: Spaced-repetition review cards for algorithm patterns, scheduled before you forget.
• Pricing