AI Application Engineer

Each step builds on the one before. Open any card for the paper and resources.

1. Retrieval basics

   5 algorithms

   BM25 Classic keyword ranking by term frequency and rarity. 2009core Sentence Embeddings (Sentence-BERT) Encodes sentences into vectors compared by cosine similarity. 2019core Dense Passage Retrieval (DPR) Dual encoders map questions and passages into one space for retrieval. 2020core HNSW (Hierarchical Navigable Small World) A layered small-world graph for millisecond approximate nearest-neighbor search. 2016core Product Quantization (PQ) Splits vectors into chunks, each coded by a small codebook, for compressed fast search. 2011core

2. RAG

   4 algorithms

   Retrieval-Augmented Generation (RAG) Retrieve relevant documents, then generate grounded answers. 2020core HyDE (Hypothetical Document Embeddings) Generate a hypothetical answer, then retrieve real documents with it. 2022ext ColBERT (Late Interaction) Keeps per-token vectors and matches them late for precise yet fast retrieval. 2020core GraphRAG Builds a knowledge graph and community summaries to answer global questions. 2024core

3. Agents

   5 algorithms

   ReAct (Reason + Act) Interleaves reasoning, actions and observations. 2022core Function Calling / Tool Use Models emit structured function calls that follow a JSON Schema. 2023core Model Context Protocol (MCP) An open protocol connecting AI apps to external tools and data. 2024core Reflexion Agents write verbal self-reflections after failures and reuse them. 2023core Agent–Computer Interface (SWE-agent) LM-tailored command interfaces that boost automated bug fixing. 2024core

4. Going live

   6 algorithms

   PagedAttention (vLLM) Manages KV cache in pages like OS virtual memory. 2023core Continuous (Iteration-Level) Batching Re-forms the batch every decoding step so finished requests leave immediately. 2022core Speculative Decoding A small draft model proposes tokens that the big model verifies in one pass. 2022core AWQ (Activation-aware Weight Quantization) Protects the ~1% most salient weights identified by activation magnitude. 2023core LLM-as-a-Judge Strong LLMs grade or compare other models' answers. 2023core Indirect Prompt Injection Hides malicious instructions in content the model reads. 2023core

Goals, projects and self-tests that go with these algorithms.

• 1Zero to Deep Learning · From linear regression all the way to ResNet and the precursors of the Transformer — build intuition for how models are trained.24
• 2Understanding an LLM End to End · From tokenisation to the Transformer, modern architectures, pre-training and inference — every part of an LLM, in nanoGPT order.19
• 3Post-training & Reasoning Models · From instruction tuning and LoRA to RLHF and DPO, then GRPO, RLVR and DeepSeek-R1 — how ChatGPT and o1/R1 are made.19
• 4Generative Models & Frontier Research · Diffusion/flow matching, multimodality, new architectures, world models and interpretability — the most active research areas of 2024–2026.26