Investigation Hub

Signals become answers, not noise.

Lumi watches everything you run and turns what matters into focused investigations — each one evaluated, given a verdict, and ranked so your team knows what to look at first.

Confirmed or ruled out

every investigation gets a verdict

7 domains

from CUDA drivers to PDU phase balance

LIVE INVESTIGATION STATUS

Every investigation ends
with a verdict.

Not “done” — confirmed or ruled out. Knowing an alert was investigated and found
harmless is as valuable as finding the fault.

▷ investigate

Ready to run

Detected and queued. One click starts the evidence gathering.

◷ running

Gathering evidence

Querying telemetry, estate and logs across every relevant domain right now.

position

Confirmed

The issue is real, with evidence attached and a recommended action.

negative

Ruled out

Investigated and found benign. The alert closes without stealing anyone’s morning.

timed out

Inconclusive

Evidence didn’t arrive in the window — flagged for a human rather than guessed at.

aborted

Stopped

Cancelled by an operator or a guardrail. Nothing runs on silently.

INVESTIGATION WORKBENCH

Cross-domain, because failures are.

A GPU throttling isn’t a GPU problem until you’ve checked the rack inlet temperature
and the PDU phase balance. Lumi checks all of it in one pass.

●gpunode-a14-3 GPU thermal throttlingCRITICALdc-westpositive 4 domains · 11s

EVIDENCE GATHERED

▨GPU clocks throttled to 1,200 MHzcompute · nvidia-smiconfirmed

🌡Rack A14 inlet temperature 27°Cfacilities · DCIM+5°C vs A13

⚡pdu-a14-left phase imbalance 14%power · PDU telemetrycorrelated

⚙train-llama-70b holding 16 GPUs for 38hworkload · slurmcontributing

✓Driver / CUDA runtime versionscompute · host factsruled out

✓ Positive — thermal, not silicon

Fan 4 on gpunode-a14-3 is below nominal RPM while rack A14 inlet air runs 5°C warmer than its neighbour. The PDU phase imbalance is loading one side of the rack. The GPU is throttling correctly to protect itself — replacing the card would fix nothing.

▷ Rebalance pdu-a14-left▷ Replace fan 4 · parts on site▷ Drain node before service

Domains correlated in a single investigation

PRIORITIZED OPERATIONS

What to look at first,
and why.

Thirty open issues isn’t a work list, it’s a wall. Lumi ranks them by real impact — what’s customer-facing, what’s spreading, what’s about to breach.

1

Business impact

Customer-facing services outrank a lab host, whatever the alert says.

2

Trajectory

“Climbing” beats “steady” — issues getting worse move up the list.

3

Blast radius

A PDU fault threatening a whole rack outranks one warm node.

4

Confirmed first

Positive verdicts rank above unstarted alerts — evidence beats suspicion.

CAPABILITIES

What investigations give you.

◉

Automated issue detection

Continuous surfacing of critical and warning conditions across the whole estate.

⚗

Investigation workbench

Evidence, correlation and verdict on one record — with the next actions attached.

⇄

Cross-domain troubleshooting

App, database, GPU, network, storage, power and security checked in a single pass.

▲

Prioritized operations

Ranked by impact, trajectory and blast radius — not by alert timestamp.

◷

Live investigation status

Running, positive, negative, timed out or aborted — always current.

✓

Negative results kept

Ruled-out findings are recorded, so nobody re-investigates the same alert.

▷

One-click launch

Start any investigation from the queue; cost is shown before it runs.

Transparent cost

Credits and billable usage visible on the panel, so analysis never surprises you.

🔗

Estate-linked

Every finding tied to the real entity, ready to hand to a workflow or ticket.

WHY IT MATTERS

The triage hour disappears.

The expensive part of an incident isn’t the fix. It’s the forty minutes of five engineers checking five dashboards to work out whose problem it is.

TRIAGE TIME

40 min of guessing

11 seconds

Evidence gathered across domains before anyone joins a call.

FALSE ALARMS

chased anyway

ruled out

A negative verdict closes it without waking anyone.

WRONG-TEAM ESCALATION

common

rare

Cross-domain correlation names the actual owner first time.

TRIAGE TIME

40 min of guessing

11 seconds

Evidence gathered across domains before anyone joins a call.

👤

For the ITOps engineer

DAY TO DAY

✓

Know what to open first

A ranked queue instead of thirty equally red rows.

✓

Evidence before the bridge call

Turn up with a verdict, not a dashboard tour.

✓

Not your problem, provably

Cross-domain checks show it’s the PDU, not the app — with the data to prove it.

✓

Nothing gets investigated twice

Negative results stay on the record for the next person.

◈

For the organization

ON THE BALANCE SHEET

✓

Shorter outages

Time-to-cause is most of time-to-resolve, and this compresses it hardest.

✓

Fewer people per incident

One verdict replaces the five-team conference call.

✓

Facilities and IT stop arguing

Power, thermal and compute evidence sit on the same record.

✓

Predictable analysis cost

Credits are shown up front, so investigation spend stays governed.

Console figures are illustrative. Coverage depends on the sources and sites you connect.

Get started

Point it at last week’s
alert storm

We’ll show you how many were real, how many were noise, and how
long each would have taken to prove.

▷ investigate

◷ running

position

negative

timed out

aborted

1

2

3

4

Build

Orchestrate

Cowork

Resolve

About Us

About Us

Product

Coworker

Orchestrator

Product

Orchestrator

Coworker

Build

Platforms

Estate Models

Agents

Observability

Service Desk