Delta, Palantir and the CI/CD Pipeline of War

Table of Content

• Delta: Figma for Killing Tanks
• From 72 Hours to Two Minutes
• AI Inside: Avengers and Friends
• Palantir: The Other Backend You Don’t See
• Is This Benchmarked Like Coding Models?
• How They “Benchmark” on the Battlefield
• The Kill Chain as a Data Pipeline
• What Changes on the Ground
• Two Futures: GitOps or Kill Lists
• What This Means for Future Wars
• And for Developers Who “Don’t Have Time”

Delta: Figma for Killing Tanks

Somewhere in eastern Ukraine, someone taps on a map and a Russian vehicle has a very short future. That click runs on something that looks suspiciously like the tools you use at work, except the “deployment” is artillery.

Delta is Ukraine’s homegrown situational awareness and battlefield management system. Think “Google Maps + Jira + drone feeds,” except the tickets explode.

It ingests data from drones, satellites, radars, frontline troops, even vetted civilians, and pins all that on a live map: enemy positions, friendly units, gear, routes, plans.

The backend is cloud-native, the frontend runs in a browser on normal laptops, tablets, or phones. If you can open a web app, you can run a war room.

Initially rolled out around 2022, Delta is now deployed across all branches of Ukraine’s security and defense forces and interoperates with NATO standards and systems like Poland’s TOPAZ artillery fire control.

From 72 Hours to Two Minutes

Before Delta, the “find target → tell someone → shoot target” cycle could take up to 72 hours. Basically, waterfall.

With Delta integrated into drone units, artillery, navy and air force, the same loop has reportedly dropped to around two minutes for some strikes. That’s not “faster reports”, that’s an entirely different tempo of war.

Ukrainian officials credit Delta with helping destroy over 15 billion dollars’ worth of Russian equipment by making coordination and targeting faster and less dumb. When you correlate that with thousands of daily targets identified, you get the idea: the system turned the battlefield into a giant event stream.

AI Inside: Avengers and Friends

Delta isn’t just a big map. It’s turning into an AI platform.

Inside its ecosystem sits “Avengers,” an AI system that parses drone and fixed-camera video to detect and classify enemy hardware: tanks, IFVs, lighter vehicles.

Some public numbers:

• Roughly 12,000 enemy targets detected per week in video streams.
• Around 70% of visible enemy equipment automatically picked up.
• About 2.2 seconds to detect a target in a frame.

Those are classic computer-vision metrics, just with higher stakes than your average “is this a cat or a toaster” dataset.

Palantir: The Other Backend You Don’t See

Then you have Palantir — the company that really wants you to know it does “algorithmic warfare,” and unfortunately is not lying.

Palantir partnered early with Ukrainian authorities, providing software to the armed forces and for government data migration and analytics. Its Gotham and AIP platforms ingest:

• Commercial and military satellite imagery
• Drone video
• Signals intelligence
• Text reports and logistics data

On top of that, tools like MetaConstellation help query what commercial satellite data is available over a given area, then fuse it into a single operational picture.

Reports and Palantir leadership themselves say their software powers much of the targeting process in Ukraine, helping shorten the “sensor-to-shooter” chain and generate strike options that update as new data comes in.

So while Delta is the Ukrainian-built C2 layer and user interface on the ground, Palantir is often the heavy analytics engine feeding in high-value targets from a much wider sensor universe.

Is This Benchmarked Like Coding Models?

Short answer: not really. At least not in any public, “here’s the leaderboard” way.

In the civilian world, AI coding models get measured on neat benchmarks: CodeSignal tests, LeetCode-style tasks, pass@k metrics. You can tell which model solves more interview questions than junior devs.

In defense, AI evaluation is… chaotic. A study on military AI notes that defense benchmarks are around two percent of all AI evaluation efforts, and mostly narrow (cybersecurity, bio stuff). Core warfighting tasks — targeting, intel fusion, logistics, command and control — largely lack robust, shared benchmarks.

Generative and decision-support models work best when the operational situation looks like their training data. Out of distribution, performance drops fast — exactly when you need them most.

So no, there’s no “Delta v1.3 scored 92.5 on the NATO BattleEval-3000 benchmark.” Most of the real testing happens live, under fire, with humans in the loop and a lot of “does this actually help us survive the night?”

How They “Benchmark” on the Battlefield

Instead of static leaderboards, you basically get online A/B testing with explosions.

For Ukraine’s ecosystem:

• Delta’s usefulness is measured in time-to-strike, reduction in friendly-fire, and the amount of enemy kit destroyed.
• AI modules like Avengers are tuned on real drone footage, with dev teams measuring detection rates, seconds to identify, and number of missed or false targets.
• Palantir-like systems are judged on how much they shorten the kill chain and whether targets are actually where the system said they were when shells land.

In parallel, other defense AI projects use war games and simulation — think AlphaGo but for battle plans — where models propose strategies and human experts give success/fail feedback over thousands of scenarios.

It’s closer to reinforcement learning plus ops metrics than to static coding tests. Ugly, noisy, and very empirical.

The Kill Chain as a Data Pipeline

If you squint, the whole thing looks like a slightly cursed event-driven microservice architecture:

1. Sensors (drones, satellites, SIGINT) emit events.
2. Palantir/Gotham-style platforms ingest, clean, enrich, cluster.
3. Delta and similar C2 systems present a fused view to humans on tablets and laptops.
4. AI modules propose next actions: “this is probably a self-propelled gun,” “this convoy looks juicy.”
5. Humans approve, and fire missions get pushed to artillery units, FPV drone teams, or other effectors.
6. Battle damage assessment flows back into the data lake to retrain models and adjust tactics.

Congratulations, you have a full MLOps loop, except the “user feedback” is craters.

What Changes on the Ground

When your map updates in near real time and your kill chain runs in minutes, several things break:

• Massed formations become suicidal. Concentrating forces is now a giant “please bomb here” marker, because anything big and slow gets spotted and tagged quickly.
• Cheap drones become strategic. Ukraine and others lean hard into swarms of low-cost, attritable systems plus AI for tasking and triage, since humans cannot watch all the video anymore.
• Mid-level commanders get “IDE-like” assistance. They see recommended targets, risk estimates, suggested routes, not unlike code suggestions, but for moving battalions instead of braces.

Some analysts argue drones now account for the overwhelming majority of strikes in parts of the conflict, with AI-enhanced tasking said to be decisive in keeping up with that scale.

Two Futures: GitOps or Kill Lists

The same Palantir stack that fuses satellite imagery to help Ukrainians hit Russian artillery is also used elsewhere for far darker things: building targeting databases and “kill lists,” running AI-assisted strikes in other conflicts, and even managing humanitarian aid flows in ways critics say are politically weaponized.

That dual-use tension is real: the architecture is the same. Swap “allocate ambulances” with “allocate drones” and you’ve basically just changed a config file.

From a purely technical angle, though, Ukraine became a live-fire lab for what AI-augmented command-and-control looks like. Allies are already studying Delta to copy or integrate bits of it into future NATO systems.

What This Means for Future Wars

The direction of travel is pretty clear:

• Every battlefield becomes a giant graph database. Units, vehicles, buildings, routes — all nodes with constantly changing properties, queried in real time.
• Humans move up a layer. Less time on manual plotting, more time arbitrating between AI proposals, ethics, and politics — at least in the ideal case.
• Speed becomes a weapon. When your OODA loop runs at machine speed, you can out-cycle opponents stuck in radio-and-paper workflows.

The scary part: militaries are pushing decision speed up while still lacking good benchmarks for reliability, robustness, or failure modes in messy conflict environments. This is basically “move fast and break things,” except the “things” are people.

And for Developers Who “Don’t Have Time”

If you’re building AI systems, Ukraine is the nightmare production environment: hostile network conditions, incomplete data, adversarial behavior, no clean labels, insane uptime requirements. And yet the stack is weirdly familiar: browser frontends, cloud backend, microservices, CI for models, observability, interoperability standards.

The big difference is not the tech, it’s governance and evaluation. Civilian AI at least pretends to care about benchmarks and alignment before shipping. In defense, a lot of that is being improvised mid-battle, with very limited transparency and almost no public metrics.

So yes, war is quietly becoming another data problem. Delta and Palantir show what happens when you wire an entire battlefield into something that behaves like a distributed system. The question isn’t whether future wars will use AI; it’s whether anyone will manage to put guardrails and proper tests around it before the next big “production incident.”