Industrial AI you can verify before it touches operations.
Forecasting, optimisation, maintenance, and process models validated the way market infrastructure is validated: pinned inputs, replayable evaluation, cost-weighted gates, and sealed evidence that satisfies process-change control.
Engagement-led · operator data stays with the operator
In industry, a model failure has a physical bill.
- Industrial AI failures are physical: scrap, downtime, mis-hedged energy positions, unsafe conditions.
- Models are validated once, then run for years against a shifting process.
- Telemetry problems are misdiagnosed as model problems.
- Process-change control needs written evidence, and the model team has none to give.
- Vendor accuracy claims cannot be reproduced on the operator's own data.
Four places we do measurable work.
Energy forecasting & dispatch
Load, generation, and price forecasts that feed dispatch and hedging decisions. We pin the data, replay the horizon, and gate the error metrics so a model change cannot quietly degrade dispatch quality.
- Replayable backtests over pinned historical windows
- Error-budget gates per horizon and per asset class
- Divergence alerts when live error leaves the validated band
Manufacturing process intelligence
Yield, quality, and anomaly models sitting next to physical process lines, where a false negative becomes scrap and a false positive becomes an unnecessary stoppage.
- Deterministic evaluation over recorded process runs
- Cost-weighted thresholds instead of raw accuracy
- Sealed evidence per release for process-change control
Predictive maintenance
Failure-onset models whose value depends entirely on lead time and false-alarm rate. We measure both explicitly and record how each release performed.
- Lead-time distribution reported, not just hit rate
- False-alarm budget enforced as a gate
- Per-asset performance tracked release over release
Operational telemetry integrity
Before modelling anything, the signal has to be trustworthy. We validate ingest paths, clock alignment, gaps, and unit consistency, and record what was verified.
- Gap, drift, and clock-alignment checks
- Unit and schema conformance enforcement
- Provenance recorded from sensor to model input
Prior operator experience, described honestly.
A regional regulated utility running reliability and asset-performance reporting across distribution assets, with outage, work-management, and asset-health data spread over separate systems of record.
Data models and analytics built directly against the utility's operational systems: reliability metrics, asset-condition views, and performance reporting delivered into the tools planners, engineers, and operations leadership already used.
Reliability and asset-performance analytics adopted by 300+ operational users, replacing manual spreadsheet reporting cycles with a single governed source for planning and regulatory reporting inputs.
Attribution · This is prior operator experience from the founder's utility background, delivered through the Blanc Quant Services consulting heritage. It is not a BQS platform deployment, and no BQS validation or evidence tooling was used on this work.
Signal first, model second, evidence throughout.
Telemetry provenance, gaps, clock alignment, and unit consistency verified first.
Recorded operating windows replayed; current model performance pinned as a baseline.
Cost-weighted thresholds agreed with operations and enforced in the release path.
Sealed bundles per release, mapped to process-change control requirements.
This is early-stage, engagement-led work. We publish no industrial performance figures we have not measured on a named workload, and we will say so plainly when a problem is outside what we can validate.