MSBAI

08/30/2026

The hard part of space operations is not seeing more. It is deciding what to look at first.

There are tens of thousands of objects up there and not enough people to watch them. So the machine narrows, and the person decides.

Three minutes of OrbitGuard. What reaches the operator is not a feed. It is a short list with its evidence attached: what changed, how confident we are, and how far the call moved when we perturbed it. Then close approaches, ranked. The last clip is a training scenario, and to be clear about it, the system detects anomalous behavior. It does not declare intent. The operator does that, and the operator approves the maneuver.

The new National Space Transportation Policy targets more than 1,000 launches and reentries a year by 2030.

So the same system does it in the atmosphere. A rocket body reenters over the upper Midwest. GURU drives NASA's DAIDALUS, built to compare one aircraft against one aircraft, across every affected flight at once. Two held on the ground, two rerouted, recomputed once per second, 1,441 times, as the debris footprint grew.

Learned models accelerate and rank. Validated physics confirms and decides. And the system is built to refuse a case it cannot defend, because anything that cannot say no has not earned the right to say yes.

One question for anyone who has sat a watch: what is the one thing you would need to see next to an alert before you would act on it?

08/28/2026

Jim flew in from New York for this one. I drove 20 minutes two men who met in the same startup class in 2020. Standing in the same room in El Segundo yesterday on opposite ends of a very large table. The room was the 2026 space contracting executive forum hosted by Space Systems Command directorate of contracting: acquisition at the speed of relevance. The golden dome case study was the standup funding that arrived a physical year clock and a finish line that moved from September 30 to July 31. What SSC delivered against that: rapid prototyping agreements in under 30 days, competitive awards in under 75, standardized agreement terms so nobody burns a month arguing over boiler place. That is not process reform. That is a group of people deciding the process was never the constraint. The sharpest idea of the day came from the portfolio acquisition executives time to contract award is the wrong metric time -to combat capability at scale is the right one: save two months on award, deliver four years late and you have saved nothing. And the most honest one: integration is what keeps them up at night. Everyone agrees the interfaces have to be ruthlessly defended. Almost nobody funds the integration up front. That is the problem we built for at MSBAI. GURU drives the government’s own accredited engineering and mission software autonomously inside the boundary at supercomputer scale. On a phase 3 milestone this may GURU synthesized 18 flight vehicle geometries, and executed 54 CFD simulations across Frontier, Aurora, and Raider in space domain awareness, OrbitGuard screens 14,710 objects for anomalous behavior.

08/28/2026

Jim flew in from New York for this one. I drove twenty minutes.

Two men who met in the same startup class in 2020, standing in the same room in El Segundo yesterday, on opposite ends of a very large table.

The room was the 2026 Space Contracting Executive Forum, hosted by Space Systems Command - SSC's Directorate of Contracting. Theme: Acquisition at the Speed of Relevance.

The Golden Dome case study was the standout. Funding that arrived late, a fiscal year clock, and a finish line that moved from September 30 to July 31. What SSC delivered against that: rapid prototyping agreements in under 30 days, competitive awards in under 75, standardized agreement terms so nobody burns a month arguing over boilerplate. That is not process reform. That is a group of people deciding the process was never the constraint.

The sharpest idea of the day came from the portfolio acquisition executives. Time to contract award is the wrong metric. Time to combat capability at scale is the right one. Save two months on award, deliver four years late, and you have saved nothing.

And the most honest one: integration is what keeps them up at night. Everyone agrees the interfaces have to be ruthlessly defended. Almost nobody funds the integration up front.

That is the problem we build for at MSBAI.

GURU drives the government's own accredited engineering and mission software, autonomously, inside the boundary, at supercomputer scale. A hierarchical neuro-symbolic multi-agent system, if you want the paper version.

On a Phase III milestone this May, GURU synthesized 18 flight vehicle geometries and executed 54 CFD simulations across OLCF's Frontier, ALCF's Aurora and HPCMP's Raider. In space domain awareness, OrbitGuard screens 14,710 objects for anomalous behavior, 0.98 AUC. Same system. Two domains that share no hardware, no data standard and no program office.

On the AI question, our line has never moved and never will. Use the machine for the first cut. Keep a human on the decision. Learned models accelerate and rank while validated physics confirms and decides. Judging by yesterday, that line is drawn in the same place on both sides of the table.

About the photo. That is James. Allen Regenor, Col USAF(ret), now leading USSF business development at IBM. We met in 2020 in Warren Katz's Air Force Techstars cohort, almost entirely over video, in the year nobody met anybody. Six years on, one of us is inside the largest technology company on earth and the other is running a company on the two fastest machines in America. That is not nostalgia. That is what the accelerator was supposed to produce, and evidence that it worked.

He crossed a continent to be in that room. I got lucky on the commute.

Thank you to Sara Lawlyes and the SSC contracting team for putting real substance on an agenda.

If you were there: what was the one line you wrote down?

08/25/2026

08/18/2026

07/26/2026

07/26/2026

On July 14, the Department of the Navy released its Strategy to Weaponize Data and AI: a roadmap to an AI-first Fleet that can "out-learn and out-fight any adversary."

Within a week, our team answered with working code.
The Navy's new strategy frames data and AI as warfighting assets on par with weapons and munitions, and it prizes one thing above all: turning information into decisions, fast. That is exactly the problem our GURU architecture was built for. So when two Navy SBIR topics called for AI-driven maritime tracking and adaptive sensor management, we pointed at the sea what we had already proven in orbit.

The video below shows both prototypes, back to back.

First, GURU MarineGuard: 787 real vessels from public NOAA data, replayed through a cascade of learned models that forecast each ship's movement, flag deviations from its learned pattern of life, and hand analysts a ranked review queue instead of an unfiltered flood. The full stack runs on a laptop.

Second, an adaptive sensor resource manager add-on to MarineGuard: it measures each radar's marginal contribution to each track, projects the consequence of releasing a sensor task before proposing it, and then waits for the operator. Advisory by design.

Both inherit their DNA from OrbitGuard, our system watching 14,710 space objects at the SDA TAP Lab with 94 to 96 percent maneuver-detection accuracy. Same architecture, new domain, days not years.

These are prototypes, and we say so on screen. The trajectories are real. The sensor numbers are deliberately notional. No score is a threat call. Showing your assumptions is a capability, not a caveat, and national-security guidance now demands exactly that: AI that is reliable, robust, steerable, and controllable under rigorous test and evaluation.

Our doctrine was written for that bar. Learned models accelerate and rank. Validated references confirm and decide. Humans stay in command.

One architecture. Space, maritime, autonomous engineering, regulated nuclear autonomy. This is simply the latest sign of what this team fields, fast, where mistakes are not allowed.

Sailors, engineers, program folks: what mission should GURU learn next?

06/13/2026

No human touched a single mesh.
In this run, GURU Gen 2 generated 18 flight-vehicle geometries and ran 54 CFD simulations from Mach 0.3 to 6.8: meshing, converging, and adapting each one, autonomously -across Frontier, Aurora, and Raider at the same time, with no reserved queue slots.
This is a 60-second cut of the live demo. Ryland Adams narrates; AFRL's George Zagaris and I jump in. George's verdict in the room: "a game changer."
Full 12-minute walkthrough in the comments. 👇

06/07/2026

One to two weeks. That is what it can take an expert to turn a vehicle's CAD into a CFD-ready mesh. Last week our GURU AI agents did that setup work in minutes, live, across three of the nation's most powerful supercomputers at once.

From publicly released images of various high-speed vehicles, GURU synthesized 18 flight-vehicle geometries, built the meshes, then set up and ran 54 CFD simulations spanning Mach 0.3 to Mach 6, from sea level to 50 km, with solution-adaptive refinement locking onto shocks, vortices, and shock wave / boundary layer interaction mid-run. Watching it live, Air Force Research Laboratory - AFRL 's George Zagaris called it "a game changer" and "what every user dreams of."

This is the bottleneck the GAO has flagged across hypersonics programs: modeling and simulation setup consumes the scarcest resource these programs have, expert time. When AI agents carry the laborious setup, those experts are elevated to the work only they can do: aerodynamics, numerics, and validation against ground and flight test. And it runs at campaign scale, thousands of cases across full flight envelopes, which is what a faster national test cadence demands.

There is a second story here for the Department of Energy. A single autonomous workflow submitted these jobs live across OLCF Frontier, ALCF Aurora, and HPCMP Raider with no advance reservations. The queues decided where each job ran. That is the federated, multi-center ex*****on the Genesis Mission envisions, and it is how the nation will extract full value from the historic compute buildout now underway, from the leadership facilities to every new AI data center.

This week I had the honor of opening the visualization showcase at the OLCF User Summit, presenting in Oak Ridge National Laboratory's Everest facility, where a wall of pixels let the audience watch flow solutions converge and meshes adapt in real detail. The foundational research behind all of this, the multimodal representation training and the hierarchical agent architecture, was enabled by the ALCC program's leadership computing allocation. These results are the return on that investment, and I have been privileged to build on these systems since the Jaguar days in 2009. Thank you to the entire Oak Ridge Leadership Computing Facility team.

We want to aim this capability at real program needs. If your program is losing schedule to meshing and problem setup, I would like to hear about it.

05/07/2026

Space Domain Awareness has a data problem most teams won’t say out loud:
The real data is too sparse to train AI that operators can actually trust. So we stopped trying to fix the data — and started generating our own.
I gave this talk yesterday at the 2026 Department of the Air Force Modeling, Simulation & Analytics Summit in Colorado Springs — hosted by STARCOM, SAF/SA CMSO, and NTSA — on the Virtual Range Architecture we’ve built at MSBAI, and how it maps to the Digital Space Range and NSTTC vision STARCOM laid out this week.
Three design choices that matter:
1. Generate the data you don’t have. TLE and EO inputs from the Unified Data Library are uneven and gap-ridden. We pre-train on millions of synthetic maneuver scenarios in NASA’s GMAT, then fine-tune on real ops data. The same playbook Tesla uses for crash scenarios autopilot has never seen.
2. Train at the embedding level, not the data level. Joint Embedding Predictive Architecture — newer than transformers, built for time-series — compresses inputs before learning. Less noise into the weights, more semantic structure out. We’re hitting AUC 0.98 on maneuver detection across 14,710 space objects, and 94–96% classification accuracy.
3. Wrap the learned components in symbolic logic. A deterministic rules engine on top is what makes the system auditable to a Guardian or an accreditor. The LLM-only crowd cannot do this part. It’s the difference between a confident model and a defensible decision.
Running at ~2-minute end-to-end latency on 20,000+ objects, with linear JEPA training scalability to 4,000 nodes on Argonne National Laboratory’s Aurora.
Built under a CDAO contract administered by Air Force DTO, embedded with Space Systems Command at the SDA TAP Lab, and tested across HPCMP, Aurora (ANL), and Frontier (Oak Ridge Leadership Computing Facility).