Guides
Practical, engineer-written answers to the questions teams actually ask about production AI — reducing LLM costs, the Model Context Protocol, building evals, and shipping with agents. No fluff, no vendor spin.
LLM cost
How to reduce LLM API costs
Most production LLM spend is avoidable: it comes from sending the wrong model too many tokens, too many times. Cutting it is an engineering problem — routing, caching, batching, and context control — not a matter of waiting for prices to drop.
Read the guide →Why are LLMs so expensive?
LLMs are expensive because you pay per token on every call, output usually costs more than input, and agents re-read a growing context on every step — so cost compounds far faster than the work appears to grow.
Read the guide →How to reduce LLM latency
LLM latency is two separate problems — the wait for the first token and the time to stream the rest — and each has its own set of fixes.
Read the guide →Self-hosted LLMs: when it pays off, and how
Self-hosting an LLM means running open-weight models on GPUs you own or rent instead of calling a hosted API — it wins on high, steady volume or hard compliance, and loses almost everywhere else.
Read the guide →MCP
What is the Model Context Protocol (MCP)?
The Model Context Protocol (MCP) is an open standard, introduced by Anthropic in late 2024, that defines one uniform way for an LLM application to connect to external tools and data through a client-server protocol built on JSON-RPC.
Read the guide →How to build an MCP server
An MCP server is a small program that exposes your tools, data, and prompts to an AI client over a standard protocol — the fastest way to give Claude or Cursor safe, structured access to your systems.
Read the guide →LLM evals
How to build LLM evals
Evals turn 'the model feels better' into a number you can gate a deploy on — they are the test suite that makes every prompt, model, and pipeline change safe to ship.
Read the guide →What is RAG (retrieval-augmented generation)?
RAG retrieves the passages relevant to a question and puts them in the prompt, so the model answers from your data with citations instead of relying on what it happened to memorise during training.
Read the guide →LLM observability: what to instrument in production
LLM observability is the practice of recording enough about every model call that you can answer, after the fact, what it cost, how long it took, whether the output was any good, and what the model was actually shown. Standard APM answers none of those, because a call that returns HTTP 200 in 900ms can still be a complete failure.
Read the guide →Agentic coding
What is agentic coding?
Agentic coding is letting an AI coding agent plan, edit across files, run tools and tests, and iterate in a loop toward a goal — powerful for the first 80% of a change, and quietly expensive on the last 20% if you skip standards and review.
Read the guide →Is your codebase ready for AI agents?
Agents don't fail randomly on a repo — they fail predictably on the same things that slow down human engineers: fuzzy boundaries, missing tests, undocumented conventions. Readiness is just those weaknesses paid down deliberately.
Read the guide →Context engineering for coding agents
Context engineering is the work of deciding what a model sees on each turn: what goes in, in what order, and — the part teams skip — what gets left out. On coding agents it is usually the difference between an agent that lands a clean diff and one that thrashes, because almost every agent failure that looks like a reasoning failure is really a context failure.
Read the guide →How to measure whether AI coding tools are working
The metrics vendors hand you — seats active, suggestions accepted, lines generated — measure tool usage, not delivery. The honest version starts from the delivery metrics you already track, adds two or three that are specific to agent work, and accepts that the result will be a directional read rather than a controlled experiment.
Read the guide →Hiring & partners
How to hire an LLM cost optimization consultant
Hire an LLM cost consultant when your bill is large enough to matter but your team is too busy to instrument, route, and eval their way out of it themselves — and only hire one who leaves you owning the result.
Read the guide →How to choose an AI engineering consultancy
Three very different businesses all call themselves an 'AI consultancy'; only one of them writes production code you own, and telling them apart is most of the buying decision.
Read the guide →How to choose an MCP server development company
The MCP SDK is a weekend's work; the job you're actually hiring for is least-privilege auth, scoping, audit logging, and production hardening on servers that reach straight into your internal systems.
Read the guide →How these guides are organised, and when to read one
The guides are grouped by the four problems teams call us about, and the grouping is practical rather than taxonomic. LLM cost covers where production spend comes from and the levers that reduce it, including latency, which is entangled with cost more often than teams expect. MCP covers the Model Context Protocol: what it is, when a server is worth building, and how to harden one for a team rather than a laptop. LLM evals covers measuring quality and instrumenting production so you can tell whether a change helped. Agentic coding covers working with coding agents — what makes a repository suitable, what belongs in a context window, and how to tell whether the rollout is paying off.
A guide is the right format when you need to understand a problem well enough to make a decision about it. It explains the mechanism, gives the trade-offs honestly including where an approach stops working, and ends with something you can act on. If you already know what you want to do and need a sequenced plan with weeks attached, the playbooks are the better format. If something specific is broken right now and you want the cause and the fix, go to the error fixes. If you just need a term defined precisely, the glossary is faster than any of them.
What these guides deliberately avoid: benchmark tables we did not produce, claims about other companies' pricing or licensing that we cannot link to a source, and predictions about where the field goes next. Model prices and capabilities move faster than any page can track honestly, so where a number would be load-bearing we explain the mechanism that generates it instead and point at the pricing comparison, which carries its own verification date. Where we state a range from our own engagements, it is labelled as that and nothing more.
They are written by the engineers who do the work, which shows up mostly in what gets included. The sections that exist are the ones where we have watched teams get stuck: the cache that was enabled but never hitting, the eval suite nobody ran after month two, the MCP server that exposed an entire database as a single tool. Every guide also names where its own approach stops helping, because a guide that has no limitations section has usually not been used in anger.
If you are not sure where to start
- Your LLM bill is growing and nobody can explain it — start with why LLMs are expensive, then the cost reduction guide.
- You are being asked whether the codebase is ready for agents — start with the agent-readiness guide, then context engineering.
- You shipped an AI feature with no way to tell if it is any good — start with building LLM evals, then production observability.
- You want agents reaching internal systems safely — start with what MCP is, then how to build and harden a server.