Hard Deployments

The memory arithmetic that decides whether a model runs at all, the difference between throughput and latency once you are batching, and how to size a box before the customer buys it rather than after.

• ○The weights are the easy part of the memory budget18 minfree
• ○Every request you add makes the system faster and each request slower17 minsign in
• ○Hardware arrives on a procurement cycle and your estimate arrives in a meeting16 minpremium

Building for an environment with no route to the internet: vendoring everything your build silently assumes, moving an update through a one-way door, and producing the evidence a regulator will accept without ever phoning home.

• ○Your build downloads things you have never thought about17 minsign in
• ○Every release is final until the next transfer window17 minpremium
• ○An auditor is asking whether you can prove it, not whether it is true16 minpremium

Isolation you can defend to a security reviewer rather than describe in a diagram, scheduling that stops one tenant ruining the median for everyone, and a blast radius small enough that a bad hour is not a bad quarter.

• ○Isolation is a spectrum, and you must be able to name your position on it17 minsign in
• ○One customer's batch job is every other customer's outage17 minpremium
• ○Decide in advance how many customers one mistake is allowed to reach16 minpremium

Finding where the money and the milliseconds actually go when you run many deployments, pulling the three levers that move them without quietly trading away quality, and planning capacity against a price ceiling somebody else set.

• ○The invoice is one number and your system has forty places it could come from17 minpremium
• ○Serve it from cache, send less, or send it somewhere cheaper18 minpremium
• ○The contract sets the price and the price sets the architecture16 minpremium

What changes when an agent's work spans hours rather than seconds: state that survives the process dying, memory that stays useful instead of accumulating, and oversight of a fleet where the correlated failure is the one that hurts.

• ○An agent that keeps its plan in memory has a plan that ends with the process17 minpremium
• ○Everything the agent remembers was true when it was written17 minpremium
• ○Two hundred agents making the same mistake is one incident, not two hundred16 minpremium

Walking into a deployment that is already failing: forming a defensible diagnosis in 48 hours without the people who built it, proving what actually happened from data rather than from accounts of it, and a fortnight's plan that restores confidence before it restores features.

• ○You have two days, no context, and everybody has a different theory18 minfree
• ○The database remembers what the meeting does not18 minpremium
• ○Restore belief in week one and capability in week two16 minpremium

Service levels that mean something when you run twenty deployments rather than one, running an incident inside an environment you cannot reach, and deciding in advance what the system does when the model it depends on is unavailable.

• ○Twenty deployments, twenty different definitions of working16 minpremium
• ○Running an incident where every action needs somebody else's hands17 minpremium
• ○Decide now what the system does on the day the model is unavailable16 minpremium

Turning three bespoke builds into one thing worth maintaining: finding the abstraction that is really there rather than the one you hoped for, designing for variation without a configuration language nobody can read, and moving existing customers across without breaking them.

• ○Three customers doing the same thing differently is not yet a product17 minpremium
• ○Your configuration file has become a programming language nobody can debug16 minpremium
• ○The customers on the old code did not ask for your platform16 minpremium

What makes the work repeatable by people who are not you: evaluation infrastructure that serves a fleet rather than a demo, field evidence packaged so the product roadmap can act on it, and a team where the hardest engagements do not all route to one person.

• ○Twenty deployments need one harness and twenty sets of cases17 minpremium
• ○Product teams do not act on anecdotes, and the field mostly sends anecdotes16 minpremium
• ○A bottleneck you are proud of is still a bottleneck16 minpremium

One engagement that uses every module: an air-gapped, multi-tenant, already-failing deployment with a price ceiling. The deliverable is a design and a set of decisions you can defend for two hours against somebody trying to break them.

• ○One engagement with every constraint at once, and a two-hour defenceproject600 minpremium