From theory to the systems that demand real understanding.
Long-form pieces on how production-AI teams orchestrate agents, serve models, and build the stack, plus GATE DA essays that turn probability, linear algebra, DBMS, ML, and AI into durable concepts.
LangGraph is what you reach for when the agent has to remember
By mid-2026 LangGraph is the default orchestration layer for stateful, long-running, human-in-the-loop agents. Not because of the API — because of the checkpoints, the interrupts, and the fact that the state outlives the request.
The agent observability stack: LangSmith, Langfuse, Helicone, Arize
You can't ship agents without observability. The 2026 stack has settled into three layers — traces, evals, drift — and five vendors who each won by being unmistakably the best at one of them. Here's who picks which, and why.
Agents for legal, finance, healthcare — the high-stakes pattern
Three industries where hallucination is malpractice. The agent companies winning in legal, finance, and healthcare have converged on the same shape — scoped tasks, retrieval over generation, mandatory human checkpoints, audit logging by default. Here's what that pattern looks like in production at Harvey, Hebbia, Abridge, and Suki.
Why XGBoost is still winning in 2026
Twelve years after the original paper, and despite a parade of transformer-based challengers, gradient-boosted trees remain the default for tabular data at most enterprises. Here's why the boring answer keeps winning — and what TabPFN, CatBoost, and LightGBM actually changed.
MLOps platform consolidation: Databricks, Snowflake AI, SageMaker
The MLOps landscape of 2023 — a dozen point tools, three feature stores, four monitoring vendors, two training frameworks — has collapsed. By 2026 the workload lives on Databricks, Snowflake Cortex, SageMaker Unified Studio, or Vertex. Here's what each platform actually does, what got swept under, and who wins which workload.
Feature stores in 2026: Tecton, Feast, Hopsworks — death and rebirth
The feature store hype cycle went peak (2021), trough (2023, 'just use dbt'), and into a quieter rebirth as the data-for-AI layer for both classical ML and LLM agents. Tecton sold to Databricks. Feast survives as the open-source baseline. Hopsworks redefined itself as an AI lakehouse. Here's what the modern feature store actually is in 2026.
The vector database shakeout: Pinecone, Weaviate, Qdrant, Chroma, pgvector
The pure-play vector DB category isn't dead, but it's smaller than the 2023 hype suggested. Postgres ate the low end, Pinecone owns serverless ops, Qdrant is the open-source cost winner, and Weaviate's hybrid search is its real moat. Here's the 2026 picture by the numbers.
Embeddings in 2026: Voyage, Cohere v4, OpenAI text-embedding-3, mxbai, BGE
The embedding model market split clean. Proprietary models (Voyage 3, Cohere v4, OpenAI v3) own the quality top; open-source (mxbai, BGE-M3, Jina v3, E5) own cost and self-host. Matryoshka rewrote the cost economics. Here's how to actually choose.
Reranking at scale: cross-encoders, ColBERT, mxbai
Retrieve-then-rerank is the single biggest accuracy improvement in production RAG. Three approaches dominate — cross-encoders, late interaction, API rerankers. Here's what each costs, what each wins, and the latency budget that decides between them.
Distributed training: FSDP vs DeepSpeed vs Megatron in production
For serious pretraining or fine-tuning, you pick from three: FSDP2 (PyTorch native), DeepSpeed (ZeRO stages), or Megatron-LM (NVIDIA, 3D parallel). The frontier labs have made their bets — Llama 3 went Megatron, FSDP2 is the open-source default under 70B, and DeepSpeed survives where ZeRO offload is necessary. Here's how to choose.
Mixture of experts in production: Mixtral, DeepSeek, Llama 4
MoE went from research curiosity in 2021 to production default in 2025. The new frontier-model template is sparse, with 5-15% active parameters per token. Here's how the serving infrastructure caught up, what vLLM and SGLang actually do with experts, and the cost math that makes MoE the future.
Quantization in production: GPTQ, AWQ, GGUF, FP8 — what to ship
16-bit serving is dead for most production workloads. Here's how to pick between weight-only post-training quantization, activation-aware quantization, and the FP8 native path — without measuring it wrong.
ML platform build vs buy: a decision framework for 2026
Most teams default to one of three answers — Databricks, SageMaker, or 'we'll roll our own.' All three are wrong for the wrong team. Here's how to pick by team size, model variety, and latency budget — with the war stories that explain why.
Model monitoring in 2026: from accuracy to behavior drift
Classical ML monitoring tracks accuracy decay. LLM monitoring tracks something stranger — the model itself silently changing underneath you. Here's what production observability looks like when the failure modes don't fit the dashboards you built five years ago.
LLM recommenders: Netflix, Spotify, and the post-collaborative-filtering era
Recommendation systems are absorbing LLMs, but almost never the way Twitter expects. Netflix and Spotify both run hybrid stacks — classical CF and content-based ranking as the workhorse, LLMs at the edges for explanation, cold start, and conversation. Pure LLM-as-recommender doesn't survive contact with YouTube-scale latency or freshness.
Multi-hop RAG without the hallucinations
Single-hop RAG works until your users start asking questions that require synthesis. Naive multi-hop hallucinates intermediate facts. Here's the working architecture — explicit decomposition, iterative retrieval, self-verification — with the production examples from Harvey and Hebbia that actually ship it.
GraphRAG: when knowledge graphs beat vector search
Microsoft's GraphRAG paper showed that building a knowledge graph from documents and querying community summaries beats vector RAG on global questions. Two years on, the receipts are mixed — the technique works, but the indexing bill is what most teams remember.
Agentic RAG: when the agent decides whether to retrieve
Naive RAG retrieves on every query. Agentic RAG turns retrieval into a tool the agent calls when the answer isn't already obvious — and the production wins are about latency and accuracy at once, not either alone.
Long-context vs RAG: the 2026 verdict
Gemini's 2M-token context, GPT-5's million, Claude's 200K. The argument that long context would kill RAG turned out wrong — but for reasons more interesting than 'cost is high'. Here's where the equilibrium actually settled.
PDF parsing remains unsolved: LlamaParse, Reducto, Unstructured, Marker
Two years into the production RAG era, the single biggest blocker for most enterprise deployments isn't the LLM, the vector store, or the retrieval algorithm. It's whether your PDF parser got the table right. The vendors have multiplied; the problem hasn't been solved.
RAG evaluation in production: Ragas, Phoenix, Braintrust
RAG evaluation is harder than LLM evaluation. Retrieval and generation fail for different reasons and need different metrics. Two years after Ragas formalized the four-metric core, the tooling has split into three camps — here's what each one actually buys you.
Fine-tuning vs RAG: the settled debate of 2026
Three years of Twitter brawling about whether to fine-tune or retrieve has ended in a boring, useful answer: RAG for knowledge, fine-tuning for behaviour, and both when you actually need both. Here's the decision table production teams actually use.
KV cache management: paged attention, prefix caching, LMCache
The KV cache is the dominant memory bottleneck in LLM serving, and three ideas — PagedAttention, prefix caching, and cross-instance LMCache — have rewired how it's managed. Here's how each layer earns its place in production.
NVIDIA Dynamo, vLLM, SGLang: serving stacks at scale
NVIDIA's Dynamo (GTC 2025) reframes LLM serving around prefill/decode disaggregation. vLLM, SGLang and TensorRT-LLM all sit underneath. Here's how the four-layer stack actually works, what the throughput numbers really mean, and who picks which.