Articles

September 23, 2026 — 30 min read

Everyone argues about retrieval. Agents that remember the wrong thing fail on the write path — what to extract, how to shape it, when to supersede it, and when to forget it.

Agent memory fails on the write path, not the read path. Drawing on 2023–2026 research (Generative Agents, MemGPT, Mem0, Zep, MemoryAgentBench, MemoryArena) and the memory products from Anthropic, Google, AWS and LangGraph, this guide treats memory as a document-modelling problem: extraction and reconciliation, namespaces and provenance, supersession with validity windows, TTL and decay, scoped retrieval, tenancy, erasure and memory poisoning — ending with a MongoDB reference schema and indexes.

July 17, 2026 — 10 min read

Product knowledge is teachable — the way you explain it, the breadth you bring, and whether you bothered to prepare are not

How I interview pre-sales Solutions Architects: a six-area scorecard probed through scenario, experience, and direct questions, weighted into a single fit indicator. Why I score for clarity, breadth, and preparation over product depth — hiring for slope rather than y-intercept — and why structured interviews beat the warm conversation your gut prefers.

July 7, 2026 — 20 min read

Views, joins, stored procedures, ACID transactions — every relational capability maps to the document model. The real difference isn't the feature list. It's that there's more than one right way to model your data.

A guide for Oracle and PostgreSQL engineers to the document model. Every relational capability maps to a real MongoDB feature with a version and a ship date: multi-document ACID transactions (4.0/4.2), $lookup joins (3.2), standard and on-demand materialized views (3.4/4.2), $jsonSchema validation (3.6), and native resumable change streams (3.6) with no relational equivalent. Worked through a double-entry ledger — the deeper lesson is that document modelling is a design space driven by access patterns, not a single normal form.

May 13, 2026 — 14 min read

Part 8 of 8 — LLM Fundamentals

How to measure what LLMs can actually do — and what the leaderboards won't tell you

Every benchmark measures a slice of capability, and the leaderboards saturate faster than a new model can ship. This article covers the frontier benchmark set as of May 2026, why LLM-as-judge works (and where it breaks), and how to build the only evaluation that actually matters — the one for your use case.

April 30, 2026 — 22 min read

MongoDB's $graphLookup, Atlas Vector Search, and GraphRAG cover most graph workloads in production. A few genuinely need Neo4j. Here's how to tell which one you have.

A senior engineer's guide to graph workloads on MongoDB Atlas vs Neo4j. Covers $graphLookup mechanics, Atlas Vector Search, $rankFusion hybrid retrieval, GraphRAG architecture (Microsoft's 2024 paper, Atlas implementation), independent benchmarks (arXiv:2502.11371), the polyglot tax, where Neo4j genuinely wins (GDS algorithms, deep traversals, Cypher/GQL), and a workload-first decision framework.

April 6, 2026 — 14 min read

Part 7 of 8 — LLM Fundamentals

Techniques, patterns, and mental models for getting the most out of any LLM

The prompt is the interface between human intent and model behavior. This article covers the essential prompting techniques — from zero-shot and few-shot to chain-of-thought and structured output — and builds a mental model for why some prompts work and others fail.

March 30, 2026 — 14 min read

Part 6 of 8 — LLM Fundamentals

Quantization, KV cache, batching, and the engineering behind every API call

Training a model is half the battle. Serving it to millions of users at acceptable latency and cost is the other half. This article covers the inference pipeline — from quantization and KV caching to continuous batching and speculative decoding — the techniques that make LLM APIs fast and affordable.

March 23, 2026 — 14 min read

Part 5 of 8 — LLM Fundamentals

SFT, RLHF, DPO, LoRA, and the art of teaching a model to be useful

A pre-trained LLM can autocomplete text, but it will not follow your instructions, refuse harmful requests, or format its answers helpfully. Fine-tuning bridges that gap. This article covers supervised fine-tuning, RLHF, DPO, GRPO, parameter-efficient methods like LoRA, and when to fine-tune vs. when to prompt.

February 16, 2026 — 30 min read

A pragmatic guide to choosing your database in 2026 — with a four-layer decision framework, real-world case studies, and a 20-question interactive checklist

A research-backed framework for choosing between PostgreSQL, MySQL, MongoDB, MongoDB Atlas, and hyperscaler-managed databases (Aurora, AlloyDB, Cosmos DB). Covers a four-layer decision methodology, weighted scoring matrices, TCO analysis, seven anti-patterns, real-world case studies from Discord, Uber, Pinterest, Shopify, Airbnb, and Temenos, and a 20-question interactive checklist with decision hygiene guardrails.

February 13, 2026 — 14 min read

Part 4 of 8 — LLM Fundamentals

Data, objectives, and the trillion-token training runs that create general intelligence

Pre-training is the foundation of every LLM — a massive, expensive process where the model learns language from trillions of tokens. This article covers the training data pipeline, the loss function, the optimization loop, and the compute required to produce a frontier model.

February 12, 2026 — 16 min read

Part 3 of 8 — LLM Fundamentals

Attention, feed-forward layers, and the block that changed everything

The transformer is the engine inside every modern LLM. This article builds the architecture piece by piece — from self-attention to multi-head attention to the full decoder stack — with interactive diagrams that let you trace how information flows through the network.

February 11, 2026 — 12 min read

Part 2 of 8 — LLM Fundamentals

How LLMs turn raw text into numbers — and why the way you split matters

Before an LLM can process a single word, it has to convert text into tokens — subword pieces that balance vocabulary size against meaning. This article walks through character-level, word-level, and subword tokenization, shows how BPE works step by step, and explores why tokenization choices ripple through the entire system.

February 10, 2026 — 10 min read

Part 1 of 8 — LLM Fundamentals

The surprisingly simple idea behind the most powerful AI systems ever built

Every LLM — GPT, Claude, Llama — does one thing: predict the next token. This article strips away the hype and builds an intuition for what language models actually are, how they differ from traditional software, and why next-token prediction is far more powerful than it sounds.

February 8, 2026 — 14 min read

Part 3 of 3 — Vector Search Fundamentals

How retrieval-augmented generation turns your data into context an LLM can actually use

A hands-on guide to retrieval-augmented generation — from chunking and embedding to retrieval, prompt assembly, and advanced patterns like hybrid search and reranking. Interactive visuals at every step.

February 7, 2026 — 10 min read

Part 2 of 3 — Vector Search Fundamentals

A visual guide to the three distance metrics behind vector search

The same two vectors can look close or far apart depending on which distance metric you choose. This interactive guide visualises Euclidean distance, cosine similarity, and dot product — and shows you when to pick each one for vector search, RAG, and recommendations.