Independent verification for energy & battery claims — ML models, vendor datasheets, and public market telemetry. We don't build the model and don't operate the asset; we check whether the evidence supports the claim.
Because trust should be verifiable.
It is why a model reporting 98% accuracy in the lab can collapse on a cell it has never seen — the result was leakage, an inflated metric, or a curated benchmark, not prediction.
Buyers, investors, and insurers can't tell a sound model from a flattering one by reading the vendor's own numbers.
VolMax is the neutral check. We don't sell a competing model, so we have no incentive to flatter a number. Independence is the product.
We don't build your model — we check whether its numbers survive contact with data they haven't seen. Independent audits designed to catch interpretation artifacts.
Independent integrity audit of an SOH/RUL model or a vendor's accuracy claim. We check split integrity (cell-level vs cycle-level), preprocessing leakage, metric honesty (full-set error, no dropped cells), and physical consistency. Deliverable: a reproducible report where every number regenerates from one script.
We reproduce a claimed result on a clean, group-aware split and report the gap between the claimed and the verified number — and trace its cause. The honest number is the one that survives a new cell.
Verification of measured-signal pipelines (power quality, vibration, MCSA, PV) against physics and the sensor's real resolution — built on a test-covered signal-processing core.
Independent reconstruction of public market telemetry (ERCOT SCED, AEMO NEMWEB) for a specific asset prior to due diligence: dispatch history, demonstrated vs. nameplate capacity, and SoC reconciliation where semantics allow. Deliverable: a reproducible report with a hash chain and pre-registered rules. Target audience: funds, lenders' advisors, insurers.
A reproducible audit report; a plain verdict (where the claim holds, where it breaks, the corrected number); the regeneration script.
An accuracy number. We are not selling a model, so we make no "95%" promise — that is the vendor claim we exist to test. A verifier who guarantees a flattering result has the same conflict of interest as the vendor. We do not build the model under test, and we list no client outcomes we don't have — the proof is the public work below.
Every finding is produced by a strict procedure built on four non-negotiables. We apply it to our own work first.
Publicly timestamped before we look at the data.
Primary public sources, with a hash captured upon download.
Open code and data, DOI-archived for persistence.
Our own errors are dated and listed, never deleted.
The standard we hold your model to, we applied to our own first. These are public, reproducible, test-covered repositories — the credential that replaces a CV.
Primary frozen-rule verdict: 55.2% pass (Inconsistent); exploratory ≥10 MWh stratification: 81.8% (mean 0.98). Both populations labelled.
Primary frozen-rule verdict: 1.22% (mean 0.6339) across all 245 evaluable events; exploratory ≥10 MWh stratification: 1.71% (mean 0.7703). Both populations reported with labels. Verdict: Not determinable from public data — and that is the finding.
16 units evaluated; dispatch conformance, generalization gap, and FCAS auditability finding.
Three honest verdicts: verified, not verified, or not determinable from public data — the third is a finding, not a failure.
NASA PCoE + Severson/Attia, DOI-archived. Honest findings led by their limits: where early prognosis breaks, where impedance is observable vs predictive, capacity-regeneration isolated from true fade. Includes the worked example where the audit caught its own pipeline overclaiming three times. Every number regenerates from reproduce.py.
github.com/VolMax-Studio/Battery_Health_PortfolioTest-covered signal library (RMS, THD, DWT, Hilbert). The measurement layer the audits stand on.
Hierarchical DGA fault diagnosis, with tested boundaries.
PV fault detection on real NREL data + injected benchmarks, honestly framed.
The "PUE Loophole" audit: how PSU conversion losses mask real facility savings.
(+ Power Quality, NILM, MCSA, CWRU, Grid Frequency, VPP — domain breadth.)
View all on GitHub →25+ years in power electronics and field electrical work, now applying that hardware intuition to independent verification of energy ML. The combination is the point: a pure data scientist doesn't know why a converter loses efficiency at high frequency; a pure hardware engineer doesn't audit a model's train/test split.
VolMax stands at the intersection — between the measurement and the claim.
Physics doesn't lie. Sensors don't lie. Everything between is interpretation — and that's where I check.
Alongside the verification practice, VolMax runs hardware and edge-ML R&D — embedded signal processing on STM32/ESP32, and a hardware safety-interlock concept (analog-to-logic override, sub-2.5µs actuator-interrupt latency, Serbian IPO priority filed). These inform the measurement side of the audits and are in active development.
VolMax Studio Lab d.o.o.
Independent energy-ML verification · Serbia · EU/remote engagements
Email: volmax.core@gmail.com
GitHub: github.com/VolMax-Studio
LinkedIn: linkedin.com/in/ivan-nestorov-274157371
For audit enquiries, include the model type and dataset if possible.