Aviation Masters | Episode 4
There is a gap between what artificial intelligence can do today and what most piston aircraft owners believe it can do. On one side of that gap: decades of engine monitor data sitting unanalyzed on hard drives, mechanics troubleshooting by intuition and pattern recognition, and an FAA certification process that has historically treated software as a liability. On the other side, machine learning models are already capable of detecting engine anomalies, explaining probable causes, and flagging trends that no human would catch in a routine pre-flight scan.
Bridging that gap is the work Dr. John Sipple has devoted much of his career to. In Episode 4 of Aviation Masters, Mike Busch sits down with Sipple for one of the most substantive conversations the show has produced — not a theoretical overview of AI, but a ground-level examination of what machine learning is, how it works, and why it matters specifically to the owner of a Cirrus, a Cessna, a Bonanza, or a Diamond DA40.
This is a long-form episode. It runs nearly two hours, and the depth is earned. If you own a piston aircraft and have ever wondered whether technology might give you better information about your engine than your mechanic currently has, this conversation will reshape how you think about that question.
About Dr. John Sipple
John Sipple’s path to aviation AI is not a straight line, and that’s part of what makes his perspective valuable. He began as a mechanical engineer at the University of Minnesota, studied thermodynamics and fluid dynamics in Hamburg, spent seven years as a U.S. Army signal officer, then moved into Boeing’s Phantom Works, where he wrote missile tracking algorithms — Kalman filters, discriminating ballistic trajectories, analyzing terabytes of sensor data from defense programs. That work led him to machine learning before the field had its current name.
He joined Google when Sundar Pichai made the call to pivot the company toward an AI-first strategy, and he spent years applying deep learning to internal telemetry systems using TensorFlow. Today, he is a professor of machine learning at George Washington University, founder of Vyzerion AI, and — critically for this conversation — an instrument-rated pilot and owner of a Diamond DA40 NG with the Austro diesel engine.
He is also a Savvy Aviation collaborator, working alongside Mike Busch on two applied AI projects that are discussed at length in this episode.
AI, Machine Learning, and Deep Learning: Getting the Terms Right
Before the conversation turns to aircraft, Busch asks Sipple to untangle terms that are often used interchangeably but mean different things. The distinction matters for owners trying to evaluate what vendors are actually selling them.
Sipple describes AI as a cover concept — the broad idea of making machines do things we associate with human intelligence. Machine learning is the engine underneath: systems that learn patterns from data rather than following hard-coded rules. Deep learning is a subset of machine learning that emerged around 2012 when three things converged: GPU hardware capable of the required computation, large labeled datasets, and algorithmic advances that allowed neural networks to be stacked in deep architectures. That convergence is the reason AI went from a research curiosity to a practical tool in roughly a decade.
He contrasts this with older “expert systems” — rule-based approaches that tried to encode human knowledge directly as logical conditionals. Those systems worked well in tightly controlled environments but turned brittle in the real world, where the number of edge cases quickly became mathematically unmanageable. Machine learning sidesteps the problem by letting the system generalize from examples rather than rules.
Anomaly Detection: What the Engine Monitor Is Not Telling You
For piston aircraft owners who fly with engine monitors, this section of the conversation is the most immediately practical. Sipple explains how multivariate anomaly detection works — and why it is fundamentally different from the limit-based alerting most owners are familiar with.
A conventional engine monitor watches individual parameters and fires an alert when a value crosses a threshold. EGT above X. Oil temp above Y. That approach catches obvious failures, but it misses the patterns that precede them. An anomaly detection model trained on thousands of flights from the same aircraft type learns what normal looks like across dozens of parameters simultaneously — cylinder temperatures, fuel flow, manifold pressure, oil pressure, RPM — and flags deviations from that learned baseline before any single value reaches a limit.
Sipple walks through the concept of Explainable AI (XAI) as it applies here: it is not enough to detect an anomaly; a useful system has to explain which parameters are behaving abnormally and in what direction. Without that explainability layer, an alert is noise. With it, the system can hand a mechanic a description of symptoms rather than a raw data dump.
He also describes what he calls a “data funnel”—the architecture needed to make this scalable. Small anomaly-detection models run on every flight with minimal computational overhead. When they flag something, the more expensive explanation computation is activated. When a persistent pattern emerges, that summarized signal gets passed to a large language model capable of reasoning about probable root causes. The funnel converts raw telemetry into actionable language.
Project Spark: Savvy Aviation’s AI Diagnostics Initiative
Savvy Aviation has been collecting engine-monitoring data for years. The SavvyAnalysis database now contains over five million uploaded flights. Sipple and Busch have been developing what they call Project Spark — an anomaly detection system trained on that database, built to identify deviations in engine behavior that correlate with developing mechanical problems.
Busch’s vision for where this goes is direct: a small model running on the edge — meaning locally, without cloud connectivity — analyzing data in real time as the aircraft flies. When something anomalous appears, a plain-language explanation comes up on a display. No threshold alerts. No raw data columns. Just an answer to the question every owner asks: is something wrong with my engine?
Sipple confirms the technology is capable of this. The models are small enough to run without exotic hardware. The constraint is not computational. The constraint, as both men acknowledge, is the FAA.
The FAA’s Hard Line on AI: Training vs. Inference
The conversation addresses this directly and without wishful thinking. The FAA has published an AI roadmap that draws a clear distinction between “learning AI” — a model that continues to train while operating — and “learned AI” — a model trained and validated on the ground, then deployed with fixed weights.
The agency’s red line is on learning in flight. A model updating its own parameters while an aircraft is airborne is, from a certification standpoint, a system whose behavior cannot be fully characterized in advance. That is not acceptable in certificated aviation. A model that has been frozen after training and deployed as inference — essentially a static lookup function, however complex — is a different matter, and one the FAA is willing to work with.
Sipple explains what certification of an ML model actually requires: documentation of the training data, validation methodology, test scenarios, and performance boundaries. The model’s behavior has to be characterizable. Busch notes that for diagnostic purposes — anomaly detection in the background, no control inputs — the threshold for certification is lower than for systems that touch the flight envelope. A warning light and a text explanation on a screen does not require the same scrutiny as an autoland system.
The episode also explores why certificated aviation has fallen so far behind automotive in autonomous systems. Sipple’s observation is pointed: the GA fleet averages 45 years old. The FAA has to write regulations that apply equally to a 1968 Cessna 182 and a current-generation Cirrus. Automotive regulators faced no equivalent constraint. Busch adds what he calls the Cadillac Index: in 1968, a new Cessna 182 cost roughly four times what a Cadillac cost. That ratio has climbed to eleven or twelve. Affordability is keeping old iron in the air, and old iron complicates everything downstream.
Project Squawk: Training a Language Model on Aviation Maintenance Knowledge
The second Savvy initiative discussed in the episode is Project Squawk — an effort to train a large language model on Mike Busch’s books, webinars, the Ask the A&Ps podcast, and as much GA maintenance literature as the team can assemble.
The goal is a system a mechanic can query conversationally: describe a set of symptoms and receive reasoning about what might be causing them, what tests to run, and what the relevant service data says. Sipple sees this as the highest near-term value application of AI in GA maintenance, precisely because troubleshooting is where the maintenance system currently breaks down most consistently.
His diagnosis of why mechanics struggle with troubleshooting is worth hearing in full. The problem is not knowledge of how to perform maintenance tasks — most A&Ps can execute procedures competently. The problem is systematic diagnosis: identifying the root cause of a symptom before touching anything. Without that step, mechanics default to replacement sequences based on the most probable cause, resulting in the shotgun pattern Busch has written about for years. An AI system trained on high-quality troubleshooting data could interrupt that pattern before it starts.
Vyzerion AI and the Diamond DA40 Oil Pressure Problem
The origin story of Vyzerion AI is specific and sobering. When Diamond issued a mandatory service bulletin for the DA40 NG with the Austro diesel engine — disclosing microscopic manufacturing defects that reduced TBO from 1,800 to 900 hours — Sipple’s aircraft was affected. The service bulletin documented in-flight oil pressure failures, forced landings, and at least one well-known incident in North Carolina involving an in-flight engine fire. Nobody was injured in that event, but the aircraft was destroyed.
Sipple’s response was to analyze his own telemetry after every flight — pulling the data log, running anomaly detection, monitoring oil pressure trends and oscillations that might precede a failure. By his own account, he is an outlier in this practice. But the experience convinced him that this kind of post-flight analysis should not require a machine learning PhD. Vyzerion AI is his attempt to make it accessible: automated telemetry analysis for DA40 operators, with actionable alerts rather than raw data.
Busch’s observation here — that it is easier to change technology than to change human behavior — captures the practical problem. Engine monitor data collection exists. Analysis tools exist. What does not yet exist, at scale, is a system that does the work without requiring the operator to do anything differently than they already do.
What This Means for Aircraft Owners
This episode is not a product announcement. Project Spark and Project Squawk are works in progress. Vyzerion AI is a focused effort aimed at Diamond owners with a specific problem. But the conversation maps out a trajectory that every piston aircraft owner should understand, because it bears directly on the future of how maintenance decisions get made.
The core of Savvy’s philosophy has always been that maintenance decisions should be driven by data rather than convention or mechanic habit. What Sipple describes is a technical architecture that could make data-driven diagnosis the default rather than the exception — not by replacing the mechanic’s hands, but by giving the mechanic (and the owner) substantially better information before anything gets opened up.
If a system can flag an emerging oil pressure anomaly three flights before it becomes a forced landing, that is not a convenience feature. That is the difference between a squawk resolved on the ground at home base and an emergency on a cross-country. The economic case is straightforward. The safety case is more important.
For owners who use engine monitors and upload data to SavvyAnalysis, the infrastructure for this kind of analysis is already partly in place. The work Busch and Sipple are doing is aimed at making that analysis automatic, interpretable, and actionable — the “headwork” Savvy has always advocated for, now assisted by tools that can process data at a scale no human team could match.
Listen to the Full Episode
Episode 4 of Aviation Masters is available wherever you listen to podcasts and on the Savvy Aviation YouTube channel. The conversation runs close to two hours and covers considerably more ground than this summary can — including the three-layer defense architecture Sipple proposes for deploying AI safely in mission-critical systems, the multimodal future of richer diagnostics, and an extended discussion on using language models to query the NTSB accident database.
Subscribe to Aviation Masters for long-form conversations with the people shaping the future of general aviation.
Resources Mentioned in This Episode
Vyzerion AI: https://vyzerionai.com/
Savvy Aviation: savvyaviation.com
NTSB Aviation Accident Database: https://carol.ntsb.gov/