Engineering Notes
Technical articles on streaming forecasting, anomaly detection, concept drift, and adaptive AI — written by the engineers behind DriftMind.
Echo — Finding What You're Looking For
DriftMind detects unknown anomalies. Echo detects known signatures. Together they answer two different operational questions on the same stream: is this unusual? and is this the bearing-failure signature we saw last month? A streaming, multivariate, amplitude-sensitive pattern-of-interest engine — built for the operational reality that most failures are already known.
Read article →Real-Time Forecasting Without Retraining
33,000–48,000 predictions per second on a single CPU. Benchmarked against Adaptive ARIMA and Triggered Prophet across four NAB datasets — with honest results. DriftMind wins on drift-heavy data; ARIMA wins on stable data. The full trade-off surface, explained.
Read article →Why Time Series AI Breaks at Scale in Power Grids
Renewables, distributed energy resources, and EV adoption are invalidating historical baselines faster than centralised retraining pipelines can keep up. Why utility AI is hitting the same architectural wall telecom hit — and why the answer is edge-deployed, continuously adaptive intelligence.
Read article →Reference Implementation: Predictive Intelligence for RO Desalination
How DriftMind integrates with existing SCADA infrastructure — AVEVA PI System, AVEVA InTouch, Siemens WinCC — to forecast energy consumption, detect membrane fouling early, and recommend optimal setpoints. Results modelled from a 10,000 m³/day plant's historical data: 3–6% energy reduction, 10–15% fewer membrane cleanings.
Read article →Why I Spent Years Building a Forecasting Engine That Never Retrains
The real bottleneck in production forecasting is not accuracy — it's the cost and latency of learning. The personal story behind DriftMind, from telecom-scale frustration to 33,000–48,000 predictions per second on a single CPU.
Read article →The Future of Forecasting Isn't Prediction — It's Adaptation
A survey of seven adaptive forecasting strategies from recent literature (2022–2024): online learning, meta-learning, adaptive normalisation, memory-based methods, and more. Why the field is moving from static models to living systems.
Read article →A Linear-Time Alternative to t-SNE for Dimensionality Reduction
Sine Landmark Reduction (SLR) — a deterministic O(N) algorithm that embeds 9,000 datapoints into 3D space in under two seconds. No iteration, no randomness, browser-native. Powers the 3D scatter plot in CSV X-Ray.
Read article →Why Time Series AI Breaks at Scale, and Why Telecom Feels It First
The retraining loop is architecturally incompatible with networks that change continuously. Amazon discontinued Lookout for Metrics. Microsoft is retiring Azure Anomaly Detector. What comes next — and why cold-start, CPU-only execution matters.
Read article →Reflexive AI: A Canonical Definition
Most AI systems are designed to think. Reflexive AI is designed to react — adapting in real time without retraining, replay buffers, or centralised inference. A manifesto for AI that evolves with the data, not behind it.
Read article →CSV X-Ray: Anomaly Detection and Forecasting for Excel and CSV Users
Not everyone who needs anomaly detection has a Python environment. CSV X-Ray brings DriftMind's forecasting and anomaly scoring to anyone with a spreadsheet — upload a CSV, get instant multi-feature analysis with interactive 3D visualisation.
Read article →Multi-Perspective Anomaly Detection (MPAD)
Single-variable anomaly detection misses the relationships between signals. MPAD analyses interdependent time series by correlating metrics across sensors, locations, and functional groups — surfacing root causes that individual thresholds can't see.
Read article →