Running AI infrastructure has always demanded specialist knowledge. Planning a training run means translating a model size, a dataset, and a deadline into GPU counts, parallelism, memory, and storage. Operating the platform means learning its tools and chasing answers across separate dashboards. Both delay results.
Federator.ai Cortex lets teams do both in plain language, with a minimal learning curve. The Intent Compiler turns a one-line training goal into a costed resource plan that Federator.ai Cortex places across your GPUs, and the Megatron-LM integration gives engineers full control when they want it. Wingman AI, a built-in copilot, answers questions about the platform and the fleet from live data, with a screen built for each question, and hands approved fixes to Martin FSD. New teams can submit real work on day one.
Intent Compiler: From Plain-Language Goals to Costed GPU Plans
Describe the workload the way you would explain it to a colleague, for example “pretrain a 70B model on 2 trillion tokens within a month, with data kept in-country.” You do not need to pick a GPU type or count. The Intent Compiler, a natural language resource planner, picks the right tier and infrastructure, compiles a costed, resource-estimated spec, and matches it to capacity you already have. If nothing available fits, it can provision new infrastructure automatically, if you allow it. Every derived value is tagged with where it came from: stated by you, taken from live Federator.ai Cortex telemetry and history, computed by a documented rule, or a default. Any value that changes the GPU count or the cost is held for your confirmation, so nothing is a black box.
Teams choose how much of the plan to define themselves, from just a goal to the full infrastructure:
State an intent
Describe the goal in one line of plain language, in the portal or through the API, and let the Intent Compiler work out the rest. This is the fastest path for researchers who know what they want to train but not what it will take to run.
Declare a platform
Bring your own virtual cluster, runtime, and training configuration. Federator.ai Cortex compiles the pipeline and makes sure the job that runs uses exactly the resources you approved.
Reserve infrastructure
Request bare resources for a set term: a managed cluster on Kubernetes or Slurm, or raw nodes. Suited to platform teams that run their own stack, with clear lines on who controls what.
Engineers who prefer to set every parameter can use the Megatron-LM integration instead, choosing the model, parallelism, and hyperparameters directly. Either way, you never pick a GPU generation: Federator.ai Cortex runs one training job across mixed GPU generations and protects it with failure-protected checkpointing, so if a GPU starts to fail, its work is moved and the run continues.
Once a plan is confirmed, it becomes a binding resource envelope with a schedule of intent. Because Federator.ai Cortex knows both ahead of time, placement is planned rather than reactive: spatial-temporal GPU optimization packs work over time, space, thermal margin, and power, keeps racks with thin thermal margin for cooler jobs, and stays under the site power cap. Inside the envelope, Cortex adds and releases GPU resources from live telemetry as the job needs them.
What Teams Gain from the Intent Compiler
One Path for Every Skill Level
Researchers describe goals in plain language, ML engineers tune parallelism by hand, and platform teams reserve raw capacity. All three use the same standard requests and submission gate.
No Model Picks Your GPUs
A language model only reads your request. GPU counts, sizing, and the final plan come from deterministic, auditable code, so no model ever decides how many GPUs a job gets.
See the Cost First
Every compiled request shows its estimated resources and cost, and nothing runs until you confirm it. Teams see what a job will consume before a single GPU is reserved.
Wingman AI: An AI Copilot for GPU Fleet Operators
ProphetStor’s Wingman AI is an AI copilot that lets data center operators ask plain-language questions about their GPU fleet and get answers drawn from live operational data. Powered by a large language model and connected directly to Federator.ai Cortex, it acts as a semantic layer over the operational metadata the platform already collects. For each question, Wingman AI understands what you are really asking, finds the exact data that answers it, queries that data live, and responds in plain language alongside an interactive screen built for that question, such as a fleet heat map, a triage table, a causal graph, or an action card with buttons to approve, edit, or reject a fix on the spot.
New users can ask how to complete a task in Federator.ai Cortex and get step-by-step guidance instead of searching through documentation. Experienced operators can ask about the fleet itself:
“Why is zone-02 running hot?”
“What is the cost impact of the last thermal emergency?”
“Compare current fleet utilization to last month.”
“Which tenants will miss their SLA this week, and what would fix it?”
Unlike a general-purpose AI assistant, Wingman AI answers from your own data center. It calls live Federator.ai Cortex APIs, shows only what each user is allowed to see, and links every figure to the record it came from. ProphetStor filed for a patent on this generative UI method in March 2023, and it is now protected by patents including US 12,505,169 B2 and US 12,688,209 B2.
How Wingman AI Answers and Acts
Behind every answer is a capability registry. Each question Wingman AI can answer is a declared capability with its own data source, logic, and display. Wingman AI matches your question to the most specific capability and calls the live Federator.ai Cortex APIs it needs, such as telemetry per GPU, jobs and tenants, heat, billing meters, and SLA budgets. Because Federator.ai Cortex keeps all of this in one data model, a single answer can combine what used to sit on five separate dashboards.
Ask “why is zone-02 running hot, and what does waiting cost?” and Wingman AI combines heat, telemetry, and billing data with the root cause from Martin FSD to return one screen: a heat map, the cause and its evidence, the cost of waiting, and an action card. Approve the card, and Martin FSD checks five safety gates (physics, blast radius, safety, trust, and confidence) before it acts. Fixes that Martin FSD is already trusted to handle run on their own, and Wingman AI writes the post-mortem after an incident.
What Teams Gain from Wingman AI
One Question, One Screen
Operators ask the question they would otherwise chase across five dashboards and get one answer on one screen, combining telemetry, jobs, heat, billing, and SLA data.
Answers From Live Data
Every answer is built from live Federator.ai Cortex APIs at the moment you ask, limited to what each user is allowed to see.
Every Figure Traceable
Each number on the screen links to the record it came from, so operators can check an answer before acting on it.
From Answer to Fix
Approve an action card and Martin FSD carries it out after five safety gates. Trusted routine fixes run on their own.
Shorter Learning Curve
New users ask how to complete a task in Federator.ai Cortex and get step-by-step guidance instead of searching documentation.
Learns From Every Decision
Each time you approve, edit, or reject a Wingman AI action card, the choice is recorded with its evidence and teaches the platform, so its suggestions improve with use.
Frequently Asked Questions
What is the Intent Compiler in Federator.ai Cortex?
The Intent Compiler is ProphetStor’s natural language resource planner in Federator.ai Cortex. It turns a plain-language description of an AI training goal, such as model size, training data, and deadline, into a costed resource plan for the GPU capacity the user already has. It picks the GPU tier, works out the GPU count, parallelism, memory, and storage the job needs, tags where each value came from, and holds any value that changes the GPU count or cost for the user’s confirmation before the job is submitted.
How does the Intent Compiler differ from the Megatron-LM integration in Federator.ai Cortex?
The Intent Compiler starts from a goal described in plain language and derives the configuration automatically. The Megatron-LM integration is for engineers who want to set the model, parallelism, and training hyperparameters themselves. On both paths, users do not pick a GPU generation: Federator.ai Cortex runs one training job across mixed GPU generations and places it on the right GPUs.
Does a language model decide how many GPUs a job gets?
No. In the Intent Compiler, a language model only turns the user’s sentence into a structured request. Deriving values, sizing the job, and writing the final resource plan are done by deterministic, auditable code. The plan the user approves is hashed, and a different plan is refused at submission.
Does the Intent Compiler submit jobs automatically?
No. The Intent Compiler produces a costed plan that the user reviews and confirms before submission, and the job that runs matches exactly what the user approved. Users can optionally allow Federator.ai Cortex to provision new infrastructure when no existing capacity fits the request.
What happens after an Intent Compiler plan is confirmed?
The confirmed plan becomes a binding resource envelope with a schedule of intent. Because Federator.ai Cortex knows both ahead of time, it plans placement instead of reacting: spatial-temporal GPU optimization packs work over time, space, thermal margin, and power while staying under the site power cap. Inside the envelope, Cortex adds and releases GPU resources from live telemetry as the job needs them.
What AI copilot lets data center operators ask plain-language questions about their GPU fleet?
ProphetStor’s Wingman AI, part of Federator.ai Cortex, lets data center operators ask plain-language questions about their GPU fleet and get answers drawn from live operational data. Each answer comes with an interactive screen built for that question, such as a fleet heat map, a triage table, or an action card.
What is Wingman AI in Federator.ai Cortex?
Wingman AI is an AI copilot powered by a large language model and connected directly to Federator.ai Cortex. Unlike a general-purpose AI assistant, it answers from the operator’s own data center: it queries live platform data, shows only what each user is allowed to see, and links every figure to the record it came from.
What data does Wingman AI use to answer questions?
Wingman AI matches each question to a declared capability and calls live Federator.ai Cortex APIs for telemetry per GPU, jobs and tenants, heat, billing meters, and SLA budgets, limited to what each user is allowed to see. Because all of this sits in one data model, one answer can replace what used to sit on five separate dashboards. Every figure on the screen links back to the record it came from.
What is generative UI in Wingman AI?
Generative UI means the answer to a question is a screen built for that question, not a paragraph of text. When an operator asks Wingman AI something, it calls live Federator.ai Cortex APIs and generates the view that fits, such as a heat map, a triage table, an SLA report, or an action card. ProphetStor filed for a patent on this method in March 2023.
Can Wingman AI take actions in the data center?
Yes, with safeguards. When an answer calls for action, Wingman AI shows an action card with the planned change, what it could affect, and the simulated result. Approved actions go to Martin FSD, which checks five safety gates before acting, and the approval is a signed token that expires after 15 minutes. Fixes that Martin FSD is already trusted to handle run on their own, while reimaging or power-cycling a node always needs a person’s approval.
How does Wingman AI help new Federator.ai Cortex users?
Wingman AI gives step-by-step answers to questions about using Federator.ai Cortex, so new users can learn the platform while they work instead of searching through documentation. This shortens the learning curve and helps an AI data center see value from Federator.ai Cortex sooner.