Choir CorpBook an assessment
Home — Investors
Investors

Expensive assets are failing sooner than their owners can replace them.

Replacement lead times on large power equipment now run two to four years, which turns every unplanned failure into stranded capacity. We tell you what’s going to fail, when, and why — while there’s still time to do something about it. This page is the short version of what we are, what we have proven and what we have not.

The thesis

The owners are the ones carrying the risk.

Data centers, battery storage and grid-edge power electronics are being built faster than the discipline to maintain them. The equipment is expensive, the failures are expensive, and the people who own that risk currently price it on uncertainty.

What we sell

Not an algorithm — an accurate answer to when this asset degrades, why, and what it costs to wait, with a justification record attached. The decision layer is what customers buy.

Why we win technically

We match the modeling technique to the asset class and prove which one wins, rather than betting the company on a single method. That posture survives the next result whichever way it goes.

Why the record matters

Insurers, lenders and warranty providers cannot price on a black-box score. An auditable degradation record is the artefact that lets them reassess risk on their own terms.

Where we actually are

What we have proven, and what we have not.

5
Completed engineering work packages, each with a written report, an ablation and an internal review pass.
3
Asset classes tested on public or simulated data — batteries, power electronics, IGBT.
3
Published misses: cases where a simpler model beat ours, reported as found.
4
Products spanning prediction and the commercial decision layer.
Real commercial cells (MIT/Stanford) Best performer: Plain data-driven model physics did not win IGBT reproduction (NASA PCoE) Best performer: Standard LSTM physics did not win NASA battery ablation Best performer: Temperature-aware formulation physics did not win Simulated SiC devices Best performer: Neural model — the architecture, not the physics physics did not win
The technique scorecard, unedited.This is the same chart we show customers. None of the four evaluations shows a physics term improving accuracy. On the simulated SiC data we re-ran the comparison at a correct training budget and the physics advantage fell to 0.6%, not statistically significant. The neural model did beat the gradient-boosted baseline on all four measures, on both test conditions, in every one of ten runs — that is the architecture, not the physics. The held-out band was used both to select the model and to run that comparison, so treat the margin as optimistic. Figures under diligence.

Several results are strong but not yet validated on measured devices. Their figures are held out of public material until that exists, and are available under diligence with their sources attached.

Technical readiness

It runs on the hardware it ships with.

Two physics models, three asset classes, four hours of continuous concurrent inference on an NVIDIA Orin NX.

1.70 ms
p99 inference latency — battery remaining-capacity model.
0.31 ms
p99 inference latency — SiC surrogate model.
4 hours
Continuous concurrent operation, both models, no drift or intervention.
2% GPU
Peak GPU use — inference ran on the CPU, leaving the GPU idle for heavier models and higher-rate signal processing. Not headroom for more assets.

Measured on an NVIDIA Orin NX during an edge-deployment benchmark, May 2026 (SOW-EDGE-5). These are deployability measurements — latency, endurance and GPU headroom — not accuracy claims. CPU ran near its ceiling at peak, so how many assets one box can carry is a separate validation.

● Diligence posture

Every number has a receipt.

We keep a Numbers-of-Record register: each figure, the study it came from, the data behind it, and its verification status. Nothing reaches a public page before it clears. That is also how the material we send you is assembled — claims at exactly the level their source supports.

Request the investor materials →
What you will get
The register: every figure and its source of record
Study results held back from public material, with status
The published misses, and what we changed after them
Deployment model, partner stack, and what each partner owns

Materials are shared under NDA. Modeled economics are labelled as modeled, with assumptions visible.

Talk to us with the register open.

We would rather walk you through what is verified, what is pending, and what failed — than hand you a number that moves.

Book a zero-risk assessmentSee the evidence
Software-first · no raw data leaves your network · every number traces to a dated report