AI data centers face a constant stream of drift, faults, and capacity changes. Static runbooks can’t keep up, and the engineers who can are scarce: every 10 MW of GPUs needs 20 to 40 of them. But automation feels risky when one wrong action can stall a weeks-long training run, so most teams keep a human in every decision, and operations move at the pace of the on-call engineer.
Martin FSD is the autonomous operations agent in Federator.ai Cortex, ProphetStor’s full-stack AIOps platform for AI factories. “FSD” stands for Full Self-Driving, and the comparison is deliberate: like a self-driving car, Martin FSD perceives what is around it, reasons about hazards, and acts. It brings self-healing operations to the AI data center, covering incident triage, cooling, and maintenance. Martin FSD watches the entire fleet, detects when it drifts from where it should be, and stages a plan to bring it back. It acts only within the autonomy it has earned, and it earns more by proving it can be trusted, one type of action at a time.
Under 30 seconds
Time to detect
12 to 72 hours
Failure warning
8 GPU failure modes predicted ahead of time at 94% accuracy, so work moves before the fault lands.
Above 70%
Incidents fixed alone
How Martin FSD Works
See → Understand
Every signal, one record
Martin FSD reads signals across five domains: GPUs, heat, power, Kubernetes, and the network fabric, and brings them into one record. It acts on failure predictions from Federator.ai Cortex and detects a fault in under 30 seconds.
Understand → Check
Right diagnosis, right fix
When a fault appears or the fleet starts to drift, Martin FSD correlates signals across IT and OT layers into 8 known patterns, traces the root cause through a causal graph of 40+ links between signals and failure modes, and checks it against a physics model of the rack. Each cause then maps to one of 16 playbooks.
Check → Act
Five gates before acting
Before any playbook runs, five gates must pass: physics (the thermal twin says it is safe), blast radius (what the action could touch), safety (85°C and 20% flow floors), trust (shadow to full), and confidence (above 60%, or ask). If any gate fails or times out, the action stops. When unsure, Martin FSD isolates the issue and asks a person, with the evidence attached.
Act → Learn
Acts alone, escalates the rest
Martin FSD runs non-destructive actions on its own once it has earned the trust to do so, and in production it fixes above 70% of L1 and L2 incidents with no person in the loop. The rest are isolated and escalated to a person with a response deadline.
Learn → See
Verified, then trusted
After every action, Martin FSD checks that things actually improved and learns only from verified outcomes. Its trust moves through four modes (shadow, advisory, autonomy, and full), and Martin FSD is promoted only on evidence.
Autonomy Is Earned, Not Assumed
Martin FSD does not start with full control. Every new deployment begins in shadow mode, where Martin FSD observes and recommends while your team decides. Each action your team approves is recorded with its evidence. From there, Martin FSD moves through four modes (shadow, advisory, autonomy, and full), one type of action at a time. As Martin FSD builds a record of verified, successful outcomes for a specific type of action, such as draining a node or rebalancing workloads, its trust ledger moves that action type, and only that one, up to the next mode when the record supports it.
Promotion is earned only on evidence, and every action is recorded for review. Some actions never run on their own, at any mode: reimaging or power-cycling a node always requires human approval, and the highest-risk actions require sign-off from two different people.
Getting Started with Martin FSD
- Start in shadow mode. Martin FSD proposes but never acts, and it changes nothing in your stack or the facility. Its evidence packs accumulate for your team to review.
- Ask Wingman AI. Ask the question you would otherwise chase across five dashboards, and read the answer as a screen built for the incident.
- Approve the first actions yourself. Each approval is recorded with its evidence, and the trust ledger moves that action type up one step when the record supports it.
- Let autonomy take L1 and L2. Above 70% of those incidents close with no person in the loop, and Wingman AI still writes the post-mortem.
What Martin FSD Delivers
Live Command Center
One view shows what Martin FSD is handling on its own and what needs a human decision. Dashboards show what the cluster is doing; Martin FSD and Wingman AI show why, and what to do next.
Plain-Language Commands
Direct Martin FSD by typing or speaking through Wingman AI, for example “drain rack 7, migrate training jobs first.” It previews every step and what it could touch; nothing runs until the request passes the five gates and, where needed, your approval.
The Cause, Not the Alarm
A blocked coolant line and a sudden workload spike look the same on a dashboard. Martin FSD tells them apart by checking each diagnosis against a physics model of how a GPU heats as coolant flow drops.
Predictive Remediation
Martin FSD acts on Federator.ai Cortex predictions, including GPU failure prediction, which forecasts 8 failure modes 12 to 72 hours ahead at 94% accuracy, so a failing GPU’s work is checkpointed and moved off it before the fault interrupts training or inference.
Incidents Closed Without a Person
Martin FSD detects faults in under 30 seconds and, in production, resolves above 70% of L1 and L2 incidents with no person in the loop. The rest are isolated with an evidence pack for your team, and Wingman AI writes the post-mortem.
Every Action on the Record
Every approval and every action is recorded with its evidence, so your team can see why Martin FSD acted, what it touched, and how each action type earned its current mode.
Frequently Asked Questions
What is Martin FSD in Federator.ai Cortex?
Martin FSD is the autonomous operations agent in Federator.ai Cortex. It reads signals across five domains, detects and diagnoses issues, stages fixes, and carries them out within the level of autonomy it has earned. In production, it resolves above 70% of L1 and L2 incidents with no person in the loop.
What does FSD stand for in Martin FSD?
FSD stands for Full Self-Driving. Like a self-driving car, Martin FSD perceives its environment, reasons about hazards, and acts with a trust it earns over time. It runs incident triage, cooling, and maintenance for AI data centers within the autonomy it has earned.
Can Martin FSD take action without human approval?
Yes, for non-destructive actions it has earned the trust to run. Every step must first pass five gates: physics, blast radius, safety, trust, and confidence. A reimage or a power cycle always needs a person to approve, and the most dangerous actions need two different people. Commands given in the console are previewed before anything runs.
How does Martin FSD earn more autonomy?
Martin FSD earns autonomy one action type at a time, through four modes: shadow, advisory, autonomy, and full. It begins in shadow mode, where it proposes but never acts. Each action your team approves is recorded with its evidence, and the trust ledger moves that action type up one step when the record supports it. Promotion happens only on evidence, and reimaging or power-cycling a node always needs human approval at any mode.
How does Martin FSD learn?
After every action, Martin FSD checks that things actually improved, and it learns only from verified outcomes. Each action your team approves is recorded with its evidence, and that record is what moves each action type to its next mode, from shadow toward full.
How does Martin FSD find the root cause of an incident?
Martin FSD reads signals across five domains: GPUs, heat, power, Kubernetes, and the network fabric. It correlates them into 8 known patterns using patented technologies, including Multi-Layer Correlation, traces the cause through a causal graph of more than 40 links to 16 failure modes, and checks the result against a physics model of the rack. This is how it tells a blocked coolant line from a workload spike.
How is Martin FSD different from runbook automation?
Runbook automation runs a fixed script when an alert fires. Martin FSD diagnoses the cause first, checks it against a physics model of the rack, selects one of 16 playbooks, and runs it only after five safety gates pass. It learns only from verified outcomes, so each action type earns autonomy on its own record.
How do you deploy Martin FSD safely?
Start in shadow mode, where Martin FSD proposes but never acts. Review its evidence packs and approve its first actions yourself. The trust ledger promotes each action type only when the record supports it, and once an action type has earned autonomy, Martin FSD handles it on its own.
How does Martin FSD work with Wingman AI?
Martin FSD is the machine’s side of the operations loop: it detects, diagnoses, and acts. Wingman AI is the person’s side: it answers questions in plain language, shows Martin’s evidence on a screen built for the incident, and collects a signed approval, which expires after 15 minutes, when an action needs a person.
Does Martin FSD replace NVIDIA Mission Control?
No. Mission Control provisions and recovers nodes on the systems it supports, and its recovery begins when an anomaly is detected. Martin FSD, the autonomous operations agent in Federator.ai Cortex, works alongside it: Federator.ai Cortex reads Mission Control’s health and recovery events and hands it a drain request ahead of a predicted failure, so the GPU is drained before the fault rather than after.