Skip to content
A pen doodle of Louie: short spiky dark hair, clear glasses, a wide smile with stubble, a bead necklace, and a hooded sweatshirt.

Louie Sakoda

Status
Simulated data
Topics
Operator workflows, Design engineering

Neuron Shift · Independent prototype

Neuron Shift Operator Console

Neuron Shift: an independent prototype exploring asset context, reversible decisions, and a handoff that preserves the reason behind an action. Simulated data; no live model or operator research.

Problem
At shift change, operations systems keep what happened but lose why a person chose what they chose. The costliest mistake is undoing a decision a colleague made on purpose.
What I did
I designed and built an independent prototype with an AI pair: a shift brief that keeps each decision and its reason, AI that attaches to assets instead of a chat box, and friction that scales with how hard an action is to undo.
Result
A working prototype anyone can try, with simulated data and no live model. Its main open question is still untested: whether a real operator finds the six-field handoff useful.
On this page
  1. The handoff
  2. Position on chat
  3. Design decisions
  4. Color semantics
  5. The unguided visitor
  6. Making it work
  7. Limits and next test

Independent concept prototype. Not built for, affiliated with, or endorsed by any company. All data is simulated. No live model, telemetry, or operator interviews.

The Neuron Shift console. An investigation rail on the left, the power path in the middle running from the utility feed down to two racks, and the selected asset on the right. The affected path is traced in amber, and two assets are outlined in red because a decision is owed on them.
Neuron proposes, the operator decides.The console after taking the shift. Amber traces the investigated path. Red marks a decision owed.

Try the prototype(opens in a new tab) · Full case study(opens in a new tab) · Source code(opens in a new tab)

Premise

At 19:00 one operator goes home and another takes over the building. What changed, what was left open on purpose, and what must they not touch?

The gap is not data. Systems keep what happened and lose why a person chose it.

Put the right knowledge into the exact moment responsibility changes hands.

Neuron proposes, the operator decides, and the decision and its reason become the record the next shift inherits. For the arriving shift engineer, the costliest mistake is undoing a decision a colleague made on purpose.

The shift brief. Item 01, UPS-A1 battery impedance, is open and shows six rows from what changed to what you need to do, with an Open on graph button. Below it, a list of what changed since 19:02.
The shift brief opens on load and keeps the reasoning, not only the data.

Position on chat

A chat box asks for a question when the operator has the least time to form one, and its answer does not pass to the next shift. So every AI surface follows three rules:

  • It attaches to objects. Selecting an asset is the query.
  • Its output is interface state. Failure impact is drawn on the graph.
  • What you accept or override joins the record. AI reasoning becomes inheritable.

Chat still suits the question nobody anticipated. I scoped it out.

Design decisions

Autonomy is shown, not implied. Each recommendation lists what Neuron does alone, what needs approval, and what it must never touch. That promise builds more trust than a confidence score.

Friction scales with reversibility. Raising an alert threshold is one click. Transferring critical load takes four checks and the asset name typed back.

The decision dialog for transferring critical load. It lists what Neuron does alone, what needs approval, and what it never does, with the last two struck through. Approve is selected, four pre-flight checks are ticked, and UPS-A1 is typed into the confirmation field.
Friction scales with reversibility.A consequential approval. Nothing is executed; the prototype records the decision.

Approve and Reject is the wrong pair. The choices are Approve, Defer with a trigger, Approve with changes, and Reject. An override needs a reason: expert disagreement with the model is the knowledge the industry is losing.

The learning layer is not the product. Real operators know what a PDU is, so domain teaching is a separate layer, off by default.

The inherited record

  1. What changed
  2. What we know
  3. What was decided
  4. Why
  5. Who decided
  6. What you need to do
Each shift-brief item keeps six fields: what changed, what we know, what was decided, why, who decided, and what you need to do.

Color semantics

Color only means state. Green is flowing or decided. Amber is an equipment condition to watch. Red means a person owes a decision. Rack R-42 sits at 92 percent of its budget and stays amber, because nobody owes a decision on it.

The power path at rest. SWGR-A and UPS-A1 are outlined in red with Needs your decision. Rack R-42 reads 18.6 kW in amber.
Before: two decisions owed.
The same power path after both recommendations were approved. No red remains. Each asset reads Decided by you, and a green Approved card is pinned beside each one.
After: both decided. Rack R-42 stays amber.

Losing power is shown as absence, not as red. The flow stops, the line turns to gray dashes, and the asset goes dark.

Failure preview for UPS-A1. Its lines are dashed, but PDU-05 and both racks are outlined in green and labeled Held by redundancy.
If UPS-A1 fails, redundancy holds everything below it.
Failure preview for PDU-05. Both racks below it go dark with dashed outlines, struck-through readings, and the label Loses power.
If PDU-05 fails, both racks go dark.

The unguided visitor

In a portfolio, nobody walks the visitor through it. So a seven-step tour runs on the first visit.

  • Every step asks for a real action, not a description of one.
  • Steps complete on state, not on clicks. Exploring ahead is never punished.
  • "Do it for me" covers every task except the consequential approval.
  • Phones get their own page instead of a squeezed console.
The guided tour on step 5 of 7. A light card reads: now compare a single point of failure. PDU-05 is ringed on the graph, both racks below it are dark with struck-through readings, and the task Select PDU-05 on the graph is checked off.
Step 5 of 7. The card places itself where it hides the fewest assets.

Building the prototype

Impact is computed, not written. The preview removes the selected asset and reruns reachability from the utility source. The topology is simplified and does not model generators.

Accessibility is part of the work. Nodes are keyboard selectable, dialogs trap focus, and every colored state is named in text.

How it was built. I built it with an AI pair in React 19, TypeScript, Vite, and React Flow. I wrote the plans and acceptance criteria, checked every result in a real browser, and made every product and design call. The AI wrote most of the code.

Limits and next steps

Reasoning and confidence figures are labeled illustrative. Decisions persist in the browser only. I do not claim to know data center operations, only enough to build something an expert can correct.

Next:

  1. Test it with one real operator. Six fields felt right to me, but nobody who runs a facility has confirmed that.
  2. An authoring flow. How does a tired operator at 18:55 write one honest sentence about why?
  3. A timeline that shows model confidence and human decisions together.