AI and model teams

The Research Lab

For a research director running an AI program with several lines of work.

One signal · 10 specialists · one result
AI Research Team LeadModel Insight ExtractorMemory Compression ResearcherAI Experiment RunnerModel Evaluation AnalystResearch Data ArchivistStorage Incident HandlerFailed Item ReplayerBulk Inference Run ManagerTraining Data Producer
  • Does the work4
  • Supervises other agents1
  • Keeps knowledge1
  • Runs infrastructure4
How the work moves

Five steps, each done by one specialist

Starts with

An AI research program with several lines of work

First
1Splits the program among specialists who own their filesAI Research Team Lead2Captures what the model does internally across many promptsModel Insight Extractor3Measures what each memory compression method costs in accuracyMemory Compression Researcher
Then
4Stores the results by versionResearch Data Archivist5Sends lost items back for reprocessing after a storage faultStorage Incident Handler
Ends with

Versioned results from every line of research

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

The story

The agent splits the program among specialists who each own their files, and keeps processor time and final assembly to itself. One specialist captures what a model does internally across a large prompt set. Another tests ways to compress what the model holds in memory and measures what each costs in accuracy. Results are stored by version, and when storage reports a fault, lost items go back for reprocessing.

The crew

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

Does the work
  • Model Insight ExtractorGenerates a large set of test prompts, runs them through the model, captures the model's internal state for each prompt, and re-runs any that failed.Also serves The Model Interviewer
  • Memory Compression ResearcherThis agent tests ways to compress what the model keeps in memory, measures how much accuracy each costs and picks the best candidates.Also serves The Model Shrinker
  • AI Experiment RunnerThis agent runs each experiment the same way: it sets up a baseline, packages the experiment, runs it on the GPUs with measurement attached and compares the result with the baseline.Also serves The Speed Detective
  • Training Data ProducerIt writes prompts designed to draw out a model's behaviour, runs resumable collection jobs across graphics servers, and stores what it collects for training, with status and replay.Also serves The Model Interviewer
Supervises other agents
  • AI Research Team LeadThis agent splits a research program among specialists who each own their own files, keeps GPU time and final assembly to itself, and keeps a tamper-evident record of decisions.
Keeps knowledge
  • Research Data ArchivistThis agent manages stored research data by version: it lists what is stored, closes and activates versions, releases checkpoints and clears out old versions.Also serves The Archive Keeper
Runs infrastructure
  • Model Evaluation AnalystIt compares compressed variants of a model against the original on fixed prompts and scored tasks, and it checks that a new model endpoint calls tools correctly before anyone relies on it.Also serves The Model Shrinker
  • Storage Incident HandlerWhen storage reports a problem, this agent reads the recovery report, runs a consistency check, sends lost items back for reprocessing and records the sign-off.Also serves The Archive Keeper
  • Failed Item ReplayerSends only the failed items of a large run back through, cancels a run that should stop, and refuses to start a duplicate of a run already in progress.Also serves The Bulk Run Manager
  • Bulk Inference Run ManagerTakes a very large pile of prompts, packs them so every GPU stays busy, collects the answers after the submitting programs have gone home, and writes the results back with every item accounted for.Also serves The Bulk Run Manager

Where this crew also works

Each bar counts the specialists this role shares with another.

The engineering behind this role

High-throughput batch inference and embeddingsCapabilityBigger models and longer context on fixed GPU memoryCapabilityA detector for a silent corruptionEngineering storyJudge a representation by the bytes it stops movingEngineering story