Michael Gurule
Applied Scientist and ML Architect
Most of the difficulty in operational data shows up before the model does.
I build statistical models for noisy, sparsely labeled, shifting data, and design the cloud systems that keep them working after launch.
About
Principal Engineer II · Telemetry and AI Operational Intelligence
Designing the models and cloud architecture that turn operational data into early warning.
Right now I lead a greenfield anomaly detection, correlation, and forecasting platform built for a network of about 32 million broadband endpoints. Before that I spent close to a decade in consulting and applied data science across financial services, insurance risk, and industrial commodities, learning each industry well enough to understand what its data was saying.
MS Applied Mathematics · BS Finance and Risk · AWS Certified Machine Learning Engineer
Read the full storyCurrent Work
The network already reported its health every five minutes. The hard part was reading it: hand-set thresholds that couldn't tell a daily traffic peak from a fault, and alerts that could only be grouped by whichever team a device belonged to.
- ~1B series scored today
- 6–12B series in scope
- 8 network domains
It started with a single streaming app. Video is nearly fully onboarded, and the remaining domains are coming on one at a time.
Baselines that know the calendar
Telemetry swings by hour, day, month, and season, and fixed thresholds treated every peak as a fault. Each signal is now scored against a baseline that carries those cycles, so normal swings stop opening tickets.
Root cause, not team boundaries
Alerts used to be grouped only by the team a device was assigned to. Correlation now follows the network topology, tracing a degradation upstream to its source and out to the services, markets, device models, and software versions it touches.
Tuned by the pipeline, checked by people
A statistical scorer sets the floor, and an automated tuning pipeline retunes the model stack on its own. Drift alerts say when to look, and every flagged anomaly is reviewed and labeled. Those labels feed back into training.
- LiveDetection
- LiveCorrelation
- Near-term readyForecasting
- In designAgentic operations
Now defining the operating flow and guardrails for agents that act on what the platform finds, alongside security and policy review.
What I Work On
Applied Science
- Regression, splines, and generalized additive models
- Bayesian and hierarchical models
- Non-stationary time series and change-point detection
- Survival analysis, Kalman filtering, and state estimation
ML Architecture
- Streaming inference on AWS for fleet telemetry
- Root-cause correlation across network topology
- Feature budgets under compute limits
- Validation under temporal dependence
- Drift detection, automated tuning, and retraining
Direction
- Platform roadmaps from detection to agentic operations
- Evaluation design and ground truth strategy
- Release standards: SLAs, thresholds, promotion
- Guardrails for agents acting on detections
Applied across network and device telemetry, insurance risk, marketing attribution, satellite health, and multi-sensor positioning.
Technical Ecosystem
Featured Work
Systems built around sparse ground truth, non-stationary signal, and sensors that disagree.
Writing
Model breakdowns, practical guides, and how I approach problems in data.
Most of it comes back to where methods hold, where they break, and what to build around them.
Averages Are the Wrong Summary for Most Things You Measure
On heavy-tailed data the mean is neither typical nor stable. A simulation of what goes wrong, and how to choose the statistic that fits the decision.
Convexity Is What Makes a Problem Solvable, and Almost Nothing Is Convex
Most applied optimization quietly swaps the real problem for a convex one. An experiment on what that swap costs, and how to tell when it is expensive.
The Newcomb Paradox: What a Thought Experiment Can Teach Us About Trusting Models
A sixty-year-old decision puzzle about a near-perfect predictor, and what it says about the gap between predicting an outcome and acting on the prediction.
Contact
The fastest way to reach me is email. I read everything and reply within a few days.