The Engine Builder

For an engineering lead shipping an AI engine to more than one kind of graphics processor.

One signal · 10 specialists · one result

Does the work1
Keeps knowledge1
Runs infrastructure8

How the work moves

Four steps, each done by one specialist

Starts with

An AI engine bound for several processor types

First

1Takes stock of memory, model files and idle capacityGPU Capacity Inventory Clerk 2Builds the engine for each processor type and runs testsAI Engine Build and Benchmark Agent

Then

3Measures throughput and saves the resultsPerformance Benchmark Analyst 4Lets requests in only when the shared card has roomShared GPU Traffic Warden

Ends with

A tested engine for each processor, with throughput results

the signalthe resultEach step is done by one specialist on the crew

The story

The agent takes stock of each machine: memory, model files and idle capacity. It builds the engine for each processor type, runs the tests and measures throughput with the results saved. Where a small helper model shares a card with a large one, it lets requests in only when there is measured room and tells the rest when to try again.

The crew

The specialists behind this role, grouped by the kind of work they do.

Runs infrastructure

Performance Benchmark AnalystRuns the same fixed benchmark set on every change, reports the same four numbers each time, ramps concurrent users until a stop condition is hit, and says plainly whether the change helped.Also serves The Model Shrinker
Shared GPU Traffic WardenLets requests into the small model only when the card has measured free memory, and otherwise answers "busy, try again in this many seconds" instead of crashing the large model.Also serves The Studio Engineer
Model Serving EngineerBuilds a version of the model sized for each GPU card you own, measures how fast each card runs it, sets each card's limits for the scheduler, and copies the right build to every machine.Also serves The Model Launch Manager
GPU Capacity Inventory ClerkThis agent takes stock of memory, model weights and idle capacity on each machine and rolls it up into one picture for planning.Also serves The Watch Office
GPU Server OnboarderThis agent prepares a newly built AI engine to take work from the shared queue by writing and checking the settings files it needs.Also serves The Branch Office Engineer
GPU ProfilerThis agent profiles the AI engine at system, chip and memory level, finds idle time under steady load and records GPU use.Also serves The Speed Detective
Model CourierCopies each new model build to the remote machines that serve it and confirms every copy arrived complete.Also serves The Branch Office Engineer
AI Engine Build and Benchmark AgentThis agent builds the engine for each GPU type, runs its tests and measures throughput with saved results.

Does the work

Release EngineerScreens each candidate change for benchmark gaming, builds release images from pinned inputs, checks every hash, compares the candidate with the current baseline and promotes it, then tells the other teams to rebase.Also serves The Software Department

Keeps knowledge

Porting Pattern ScoutFinds how a reference library or specification solves a problem and where the team's own code would host the equivalent.Also serves The Migration Planner

Where this crew also works

The Branch Office Engineer22

Each bar counts the specialists this role shares with another.

The engineering behind this role

Custom small models for tool calling and routingCapability Budgets and refusalsIdea A router beside a production modelEngineering story Bringing a serving stack up on new siliconEngineering story Engines for the card you haveEngineering story All capabilitiesThe engineering, by what it does