Notes from the lab.
Working notes on RAG architecture, LLM fine-tuning, transformer-based recommendation, and the practical edges of taking AI systems to production.
Retrieval, Ranking & Recommendation
The full stack of modern recommender systems — features, retrieval, ranking, serving, and closing the loop. With code, math, diagrams, and production notes. →
-
24 — System Design Capstone: An End-to-End Modern Recommender
-
23 — Closing the Loop: Feedback, Drift, and Continual Learning
-
22 — Feature Stores & ML Infrastructure for RecSys
-
21 — Serving Architectures: Batch, Real-Time, and Streaming
-
20 — A/B Testing and Online Experimentation for RecSys
-
19 — Diversity, Fairness, and Serendipity
-
18 — The Cold Start Problem
-
17 — Position Bias, Click Models & Counterfactual Evaluation
-
16 — Vector Search & Indexing: HNSW, FAISS, ScaNN, Annoy
-
15 — Generative AI in Recommendation: LLMs, Semantic IDs, and TIGER
-
14 — Reinforcement Learning in Recommender Systems
-
13 — Multi-Armed Bandits in Recommendation
-
12 — Multi-Task Recommendation: MMoE, PLE, ESMM
-
11 — Learning-to-Rank: LambdaRank, LightGBM, XGBoost, ListNet
-
10 — Graph-Based Recommendation: GNNs, GraphSAGE, PinSage, LightGCN
-
09 — Two-Tower Retrieval: The Workhorse of Modern Candidate Generation
-
08 — Sequential & Session-Based Recommendation: GRU4Rec, SASRec, BERT4Rec
-
07 — Advanced Recommender Algorithms: DIN, DIEN, DCN, xDeepFM, AutoInt
-
06 — Neural Network–Based Recommender Algorithms
-
05 — Embeddings: From Word2Vec to item2vec and Beyond
-
04 — Traditional Recommendation Algorithms
-
03 — Target Variables: Explicit vs Implicit Feedback
-
02 — Feature Types in Recommender Systems
-
01 — Evaluation Metrics for Retrieval, Ranking, and Recommendation
-
LLM Finetuning - RLHF (PPO) vs DPO
-
Hallucinations in LLMs — Causes, Mitigation, and Deployments
-
Generative AI in Computer Vision Applications