ALICE

Meet the Team

Alex Daoud

Theo Berk

Merek Soriano

Jared Viani

Xavier Pazos

The Growth of US Data Centers

6.7-12%

4.4%

325-580 TWh

MONTH YEAR

Predicted share of US Electricity usage (by 2028)

Of US Electricity(2023)

Projected annual use by 2028

Recent Failure Cases

August 2023 – Microsoft Australia

Feb 2026 — Microsoft Azure

MONTH YEAR

July 2026 — Google Cloud

A momentary power disturbance became multi day service recovery

An automated security workflow error disabled key storage accounts required for function

A power failure followed by a fueling failure: google cloud down for a combined 15 hours

Where ALICE fits in

Observing Locally

Acting With Oversight

Optimization

Operates Offline

Reads sensors in real time.

Can perform routine functions automatically, and requests human approval for risky actions

Carefully balances critical functions within available energy levels.

Operates locally, off of the cloud and any network

What ALICE Is

A Local Decision Point

A Behavioral Baseline

A Human In The Loop

A Signed Evidence Ledger

• runs on a Raspberry Pi beside the equipment, so every requested action is judged on site instead of in the cloud.

• scores each request against what normal has actually looked like here, not only against what the rulebook permits.

• pulls in an identity-verified technician only for the requests that look wrong, and gives them the evidence to decide in seconds.

• hash-chains every request, decision and outcome to local storage, then uploads it intact once the network returns.

How a Request Is Judged

1

2

3

4

An agent does not have to be compromised, or break a single rule, to do the wrong thing at the wrong moment. So ALICE asks two questions about every action: is it permitted, and is it normal?

A signed request from a known agent

Checked against permissions held offline

Scored against the learned normal pattern

Recorded before anything is allowed to move

Four Possible Outcomes

Allow

Request Context

Deny

Hold

• the action is permitted and fits the pattern, so it executes locally and the record is written.

• something needed to judge it is missing, so ALICE asks for that context before resolving the action.

• the action is prohibited outright. No anomaly score, no explanation and no approval can override that.

• no rule was broken, but the action does not fit normal here, so it waits for a person.

An Example Walkthrough

1

2

3

4

5

6

The site loses connectivity.

A bay overheats and the cooling agent acts

Rising fan speed pushes power past a limit

ALICE provides context to the operator, and they make the decision

Agent conflict requires human in the loop

A second agent, asks for an immediate drop to zero to pull power back down.

Introducing Our Project: A demo

Expanding into Industrial Systems

Most of our country’s critical infrastructure follows the same operational pattern: observe local conditions, predict changing demand, coordinate equipment within safety limits, and continue operating when connectivity fails.

A REPEATABLE DEPLOYMENT PACKAGE

1

Sensor/Actuator

Mappings

2

Operating parameters

and equipment limits

3

Approval by a human

for low confidence changes

4

A local behavioral model

and evidence ledger

The operating pattern repeats. Each site changes the sensors, limits, and approval rules.

Our Next Pivot

Cold Storage and Food Processing

Battery Management

Mining Ventilation

Water and Wastewater Systems

MONTH YEAR

• preserve temperature using compressors, fans, door sensors, and backup power when connectivity fails.

• prioritize critical loads across solar, batteries, generators, and controllable equipment.

• control fans from gas, dust, temperature, and occupancy sensors while protecting both workers and the facility’s energy reserve.

• maintain pressure, tank level, and minimum flow while shifting pumps away from peak demand.