Practice

Implement the primitives that ML engineering interviews and real systems are built on. Every problem runs in your browser. No setup, no submission queue — just write Python and press Run.

440 problems29 topics35 companies

Featured

Prep by company

Google

Search, Ads, DeepMind, YouTube. Strong on classical ML, ranking, and large-scale serving.

107 problemsStart prep →

Meta

Meta AI / Reality Labs / Ads ranking. Heavy on PyTorch internals, recommendation systems, multi-task learning.

88 problemsStart prep →

OpenAI

LLM training and post-training. Expect transformer internals, attention variants, sampling, and tokenization.

57 problemsStart prep →

Anthropic

Frontier LLMs and safety. Transformer mechanics, RLHF/RLAIF, evals, and inference-time techniques.

55 problemsStart prep →

Difficulty:Topic:Company:

0/440 Solved

#	Title	Difficulty	Topics	Companies
1	Accuracy score	Easy	Evaluation Metrics
2	AdaBoost fit (decision stumps)	Medium	Ensemble LearningBoosting	Zomato
3	Adadelta optimizer	Medium	Calculus
4	Adagrad optimizer	Easy	Calculus
5	Adam optimiser step	Medium	Calculus	Meta OpenAI Anthropic Zomato
6	Adamax optimizer	Medium	Calculus
7	AdamW step (decoupled weight decay)	Medium	CalculusRegularization	Zomato
8	Affine coupling layer (normalizing flow)	Medium	Neural Networks
9	Assemble the Critic's audit input	Easy	Agentic AI
10	Build the chat messages list	Easy	LLMsAgentic AI
11	Cache-before-search, write-after-fetch	Medium	Agentic AI
12	Which ReAct step are we on?	Easy	Agentic AI
13	Count the dedup savings	Easy	Agentic AI
14	Find the bottleneck executor	Medium	ML System DesignAgentic AI
15	Format retrieved docs as context	Easy	LLMsML System DesignAgentic AI
16	Rebuild findings from gather results	Easy	Agentic AI
17	One model, three roles — route by prompt	Easy	LLMsAgentic AI
18	Accumulate findings across attempts	Easy	Agentic AI
19	MockLLM keyword routing	Medium	LLMsAgentic AI
20	MockSearchEngine keyword retrieval	Easy	LLMsML System DesignAgentic AI
21	The Orchestrator feedback loop	Medium	Agentic AI
22	Parallel execution profile	Medium	Agentic AI
23	Parse the planner's sub-question list	Easy	Agentic AI
24	Parse the SEARCH / ANSWER protocol	Easy	Agentic AI
25	Parse the Critic's structured verdict	Medium	Agentic AI
26	Plan → execute → synthesize	Medium	Agentic AI	Deloitte Salesforce
27	The RAG run: search → ground → answer	Medium	LLMsML System DesignAgentic AI	Accenture Amazon Deloitte IDFC First Bank Microsoft Zomato
28	The ReAct loop with a max_steps guard	Medium	LLMsAgentic AI	Deloitte Salesforce
29	The whole system, expressed	Medium	Agentic AI
30	The SharedMemory store	Easy	Agentic AI
31	Aggregate trace events	Easy	ML System DesignAgentic AI
32	Run totals from the trace	Easy	ML System DesignAgentic AI
33	Validate a chat message	Easy	LLMsEvaluation MetricsAgentic AI
34	Sequential vs parallel wall-clock	Easy	Agentic AI
35	Add & Norm (post-norm residual)	Medium	TransformersNormalization
36	Fuse token + positional embeddings	Easy	TransformersEmbeddings
37	Attention step 3 — weighted sum of values	Easy	Transformers
38	Teacher forcing — next-token targets	Medium	TransformersNLPNeural NetworksLLMs
39	Concatenate the heads	Easy	Transformers
40	Cross-attention — decoder Q, encoder K/V	Medium	Transformers	Adobe
41	Cross-attention score shape (target × source)	Easy	Transformers
42	One decoder layer (post-norm, 3 sublayers)	Medium	Transformers
43	Head dimension d_k = d_model / h	Easy	Transformers
44	The input layer — token lookup + positional	Medium	TransformersEmbeddings
45	Per-layer complexity — attention vs recurrence	Medium	TransformersRNN
46	Map a model to its transformer family	Easy	TransformersNLP	Accenture Adobe ServiceNow
47	Why total compute is invariant to h	Medium	Transformers
48	Why position matters — permutation-invariant pooling	Easy	Transformers
49	The position-wise feed-forward network	Easy	Transformers
50	Project X into Q, K, V	Easy	Transformers	Adobe

Showing 1–50 of 440