Everything, in one place.
One curriculum for the principal engineer loop with AI: 13 phases that pull together the posts, interview rounds, deep dives, reels and revision cards. Follow the phases in order or pick a path.
Learning paths
each step says why it comes nextCrack AI interviews
Four connected paths: AI interview hacks, every new round, a study plan with a final stretch with AI off, and a mock interviewer you build yourself. Plus critical basics to memorise.
React Native too: every React Native page in one place.
Mobile principal in two weeks
Follow Priya's edit from a tap in a basement to her supervisor's tablet, and build each part as you meet it. Each day changes one constraint. You will notice how much you already know.
- 1 Basement
- 2 Save
- 3 Wake-up
- 4 Sync run
- 5 Collision
- 6 Server
- 7 Big files
- 8 Mock one
- 9 Screen
- 10 Chat
- 11 Payments
- 12 Feed
- 13 Platform
- 14 The loop
AI engineering track
From how a model works to shipping LLM features.
Curriculum
Phase 00 · 0/2FoundationsWhat the OS does to your app and why offline-first is the answer.Not started
Phase 01 · 0/5Local data and the sync protocolSQLite, the outbox, cursors, idempotency and conflicts.Not started
Phase 02 · 0/6Files and background executionResumable transfers and what Android and iOS allow in the background.Not started
Phase 03 · 0/3Front end and React NativeUI architecture, sync-aware screens and the native split.Not started
Phase 04 · 0/4Backend at scaleChange logs, CDC, snapshots, fan-out and protecting the API.Not started
Phase 05 · 0/6Mobile system design roundsSix full rounds: offline, chat, feed, payments, booking, release.Not started
Phase 06 · 0/5AI foundationsLLM internals, inference acceleration, embeddings, adaptation.Not started
Phase 07 · 0/8AI systemsAgents and MCP, integration, team acceleration, and five AI rounds.Not started
Phase 08 · 0/4Interview masteryHow panels score, decision stories, rapid revision.Not started
Phase 09 · 0/29Coding roundFifteen patterns, 45 worked problems, mobile machine coding in Kotlin, and the hacks that save minutes.Not started
Phase 10 · 0/3Behavioral roundProblem-solving questions answered in STAR form, with probes and red flags.Not started
Phase 11 · 0/14Protocol ExplorerFourteen protocols decoded packet by packet: DHCP to SSH.Not started
Phase 12 · 0/51AI systems in depthServing, retrieval, fine-tuning, evals, safety and GPU infrastructure, whole to parts.Not started
Glossary
42 terms- Outbox
- Table of pending local writes, committed in the same transaction as the data, drained by the sync engine.
- Idempotency key
- A client-generated ID (op_id) that makes a retried request safe because the server applies it once.
- Cursor
- The last server sequence number a client has seen; pulls ask for changes after it.
- Tombstone
- A marker row that records a deletion so other devices can delete too.
- 410 Gone rebuild
- Server answer when a cursor is older than the retained log, telling the client to reload from a snapshot.
- WAL
- Write-ahead log mode in SQLite: readers do not block the single writer.
- Keyset paging
- Paging by the last seen key instead of OFFSET, so every page costs the same.
- FTS5
- SQLite full-text search extension using an inverted index.
- HLC
- Hybrid logical clock: wall time plus counter that never goes backwards and tracks real time.
- CRDT
- Data type whose replicas always converge, whatever order updates arrive in.
- Optimistic concurrency
- Send the version you edited from; the server rejects with 409 if it is stale.
- Multipart upload
- Splitting a file into independently uploaded and retriable parts, assembled by the store.
- Presigned URL
- A time-limited URL that lets a client talk to object storage directly.
- Range request
- HTTP request for part of a file; the basis of resumable downloads.
- WorkManager
- Android library for deferrable, guaranteed background work.
- BGTaskScheduler
- iOS framework that runs background tasks at times the system chooses.
- Background URLSession
- iOS transfer session performed by a system daemon and surviving app termination.
- Turbo Module
- React Native native module using JSI with a typed spec.
- KMP
- Kotlin Multiplatform: share Kotlin logic between Android and iOS.
- CDC
- Change data capture: streaming committed database changes from the log into a queue.
- Thundering herd
- Many clients retrying at once after an outage; answered with jitter and Retry-After.
- KV cache
- Stored attention keys and values per token during generation; it limits concurrent requests per GPU.
- Continuous batching
- Serving technique that adds and removes requests from a running batch every step.
- PagedAttention
- Managing the KV cache in blocks to cut memory waste.
- Quantisation
- Storing weights or activations at lower precision to save memory and speed up inference.
- Embedding
- A vector representing meaning so similar items are close together.
- RAG
- Retrieval-augmented generation: fetch relevant documents and give them to the model.
- Reranker
- A model that reads query and candidate together to order results precisely.
- LoRA
- Fine-tuning by training small adapter matrices instead of all weights.
- MCP
- Model Context Protocol: a standard for exposing tools and data to AI applications.
- Prompt injection
- Instructions hidden in content that try to steer a model into unintended actions.
- HNSW
- Layered small-world graphs that answer nearest-neighbour queries in milliseconds.
- BM25
- Keyword ranking that rewards rare terms, saturates repeats and discounts long documents.
- RRF
- Reciprocal rank fusion: merges rankings using positions, not scores.
- NF4
- 4-bit format whose levels follow a normal distribution, used by QLoRA.
- DPO
- Direct preference optimisation: trains on preference pairs with no reward model.
- ZeRO
- DeepSpeed's sharding of optimiser state, gradients and weights across GPUs.
- FSDP
- PyTorch's native sharded training, wrapped per transformer block.
- Triton
- Python-embedded language and JIT compiler for GPU kernels.
- NCCL
- NVIDIA's library for moving tensors between GPUs and servers.
- GQA
- Grouped-query attention: many query heads share fewer key-value heads, shrinking the KV cache.
- Eval gate
- An automated quality check that a prompt or model change must pass before release.
How each lesson works
the loopThe problem, the analogy and the mechanism, in plain words.
Diagrams for architecture and flow that you can redraw from memory.
Practice mode: write your answer, then compare with the model one.
Critical incident questions and flashcards, tracked in this browser.