Research
Edge AI, mesh, and the future of tactical ISR.
Analysis, whitepapers, and technical research on the operational, architectural, and acquisition questions behind mission software for the tactical edge.
Why the Edge Wins
Enterprise C2 consolidated behind a $20 billion ceiling and the transports settled at half a million users. Every entity those platforms fuse is produced somewhere else, during missions, at the edge. The oldest pattern in computing says the layer where data is born ends up owning the stack.
The mission application layer
Enterprise C2 consolidated above. Transports settled below. Between them sits the layer where missions actually run, and Ukraine's volunteer developers proved it exists. This is the architecture of that layer as EdgeLance built it.
Operators keep building the mission layer by hand. We shipped it as a product.
A volunteer-built artillery app reached 90-95% of Ukrainian units while the official program died in staffing. The lesson is bigger than one app. Between enterprise C2 and the radios sits the layer where missions actually run, and operators keep building it by hand. That layer has a name now.
A board asked why the operator overrode the AI. The ledger already knew.
Six months after a mission a review board asks what the AI recommended at 0347 and why the operator overrode it. In most programs that question starts weeks of forensic reconstruction. In EdgeLance it is one query against a signed append-only ledger that holds every inference, every denial and every handoff. The board adjourned before lunch.
NATO found the money at Ankara. Fielding is the hard part.
The Ankara summit closed with $139B in new allied defense investment, $50B in procurement deals in one day, and a $40B Drone Edge pledge. Rutte says the focus has shifted from setting targets to delivering results. Delivery is a software problem NATO has not solved.
Deployment complete. Now make it fight.
The install finishes at 0700 in a room with no internet. The screen shows a green check and the words deployment complete. Everyone upstream did their job, and the pipeline's job ends right there. Defense Unicorns solved delivery into air-gapped environments. EdgeLance is what the operator opens next, and it arrives inside the same signed package.
Counter-swarm is an assignment problem. Solve it decentralized or lose.
Twenty interceptors, forty inbound tracks, jammed comms, and a shrinking magazine. Which interceptor takes which track is a math problem, and the centralized solver dies with the link. We built a simulator to test distributed auction assignment under four comms conditions. Here is what it shows.
NORTHCOM defeats one in four drones detected over US bases
Waves of 12-15 drones over Barksdale's flight line in March. NORTHCOM's commander told Congress we defeat about a quarter of what we detect. Authorities have caught up. Capability has not. Base defense is now a CONUS mission, and it is an edge AI problem.
DDIL by design: an architecture for operating through denied, degraded, intermittent, and limited connectivity
Russian jamming dropped Excalibur's hit rate from roughly 70% to under 10%. The round did not fail. The link did. This paper walks through the full EdgeLance DDIL architecture. A signed append-only mission ledger with deterministic merge. Nine interchangeable transports with payload shaping by link health. Local-first AI with signed provenance. Air-gapped fleet delivery. And the acceptance tests that prove it.
The interceptor cost curve finally flipped. Software decides if it stays flipped.
Ukraine destroyed nearly 5,000 Shaheds in June. Interceptor drones took 55% of the kills at $15K per shot against a $30-50K target. The Army bought 13,000 and opened a program for more. Cheap rounds solve the ammo problem. They do not solve the assignment problem.
Warhacker outbrief: demo, feedback, ship
Four days in San Diego with a live fleet on a folding table. We demoed to Defense Unicorns and senior military leaders, took hard feedback, and the feedback is already shipping as new OS capabilities. Build-first events work. We will be at the next one.
The camera cried wolf all night. The radio knew the truck.
At 0214 the north fence camera trips on a deer for the second time and the operator stops looking. One sensor gives you an alert. Six modalities fused at the node give you a story. EdgeLance builds that story on the team's own hardware, and a $30 radio dongle that decodes tire pressure broadcasts turns any passing vehicle into a contact with a history.
We're at Warhacker. Here's what we're demoing.
EdgeLance at Defense Unicorns Warhacker, June 16-19 in San Diego. Live fleet: laptop command node, phones, a smartwatch, cameras. Mission pack to local AI to ATAK publish to mesh sync to burn. Come find us.
Edge AI hardware in 2026: laptop, tablet, and base-station silicon compared for tactical inference
Three silicon classes define edge AI in 2026. Laptop-class chips at 38 TOPS for the carry-on command node. Tablet-class at 80 TOPS for ruggedized field devices. Base-station modules at 2,070 FP4 TFLOPS. This paper compares memory budgets, thermal profiles, model fit, and practical loadout planning by device class.
The 12B multimodal release and the end of single-purpose edge models
A major lab released a 12B open-weight model on June 3, 2026 that handles text, image, and audio in a single encoder-free architecture and fits in 8GB at 4-bit quantization, under Apache 2.0. This paper covers the architecture, memory impact across device tiers, and what it changes for tactical mission loadouts.
TAK 5.6, 5,000 COTAK users, and the missing AI layer
TAK 5.6 shipped. COTAK hit 5,000 users. International adoption is accelerating. But TAK sees everything and understands nothing. The AI layer that processes sensor data and publishes intelligence as CoT already exists.
The open-weight inflection: Apache 2.0 models and defense AI without vendor kill switches
The Pentagon banned Anthropic from Maven and gave Palantir 180 days to rip it out. Open-weight models under Apache 2.0 cannot be revoked. Once downloaded, they run on your hardware regardless of what the vendor decides.
WWDC 2026: Apple just built the on-device AI framework the military needs
Core AI replaces Core ML. Third-party models plug in natively. MCP goes platform-wide. Apple did not build this for the military, but the framework solves the exact problems tactical edge AI has been working around for two years.
Hormuz and the 95% drop: what a real strait closure means for disconnected operations
The Strait of Hormuz has been functionally closed since February 2026. Transit dropped to 5% of normal volume. US forces are in active combat with kinetic and electronic denial layered on top of each other. This paper examines what the Hormuz crisis means for edge AI architecture, maritime mesh, and classification-aware data handling in a real shooting war.
UDS Fleet gets software to the device. Then what?
Defense Unicorns launched UDS Fleet to deliver software to tactical devices in classified and air-gapped environments. Delivery is half the problem. The other half is what runs after the software arrives.
Ukraine's drone mesh war is an edge AI problem
Both sides are building AI-enabled mesh for drone swarms because EW makes remote control unreliable. The arms race driving autonomous drone warfare is the same architecture problem that defines disconnected ground operations.
Meta Display glasses and the case for consumer AR at the tactical edge
$799 gets you a heads-up display readable in sunlight, a 12MP camera, and an EMG wristband for silent input. Military HUDs cost $22,000 and take years to field. Consumer AR just crossed the tactical threshold.
Maven hits $1.5B. The operator still can't use it offline.
Maven Smart System became a program of record in March 2026. Budget scaled from $103M to $1.5B requested. All of it serves combatant commands and operations centers. The dismounted operator with a degraded link gets nothing.
Lattice + EdgeLance: why enterprise C2 and tactical edge AI are better together
Lattice gives command the operational picture. EdgeLance gives the operator mission intelligence. A bridge between them lets the team decide what flows upward, covering the full stack from satellite to soldier without forcing either side to compromise.
The best training data in defense is shrink-wrapped on a pallet
A pallet of after-action binders ships to a records depot and teaches no one. EdgeLance distills those same records into unit-scoped adapters, signed and versioned fine-tuning layers that ride on any open-weight base model. Every mission the unit runs makes its model better, and the unit owns every turn of the wheel.
Ephemeral missions: how disposable mission data restores operator trust without compromising intelligence
Every device an operator carries is a potential discovery item. Enterprise platforms treat total data capture as a feature. Operators treat it as a liability. Ephemeral missions flip the default: isolated encrypted containers, selective AAR export, cryptographic destruction via hardware secure enclave, and destruction receipts. The team gets the intelligence. The raw data does not survive.
What happens when you wipe: cryptographic mission destruction and why operators will not use anything less
Ephemeral missions are not a feature toggle. They are a cryptographic architecture: destroy one key and every detection, transcription, and recommendation becomes mathematically unrecoverable across every node in the mesh. No forensic recovery, no enterprise copy.
The answer came back with two debriefs attached
Mid-mission an operator asks what vehicle was at the compound during the last two visits and gets the answer with two cited debriefs and a timestamp. A cloud chatbot would have guessed. EdgeLance mission RAG answers only from the mission vault, cites every claim and refuses when the record is silent.
After Anthropic: why sovereign AI means local AI
A vendor in California can revoke AI access to military systems based on its own policy decisions. That makes the architecture the vulnerability. Renting AI capability via API and owning it on your hardware are two different risk postures, and only one of them you actually control.
Field biometrics and AI-powered mission readiness: wearable data as a force multiplier
Every operator wears a WHOOP, Garmin, or smartwatch. The data sits on their phone and goes nowhere useful. This paper covers how to aggregate wearable biometrics, score readiness with on-device AI, and publish the result in ATAK before step-off. No data leaves the network. The medic gets individual status. Command gets a quantifiable go/no-go.
The $5,000 ISR stack: what a laptop, four phones, four cameras, and a LoRa radio can do
A flagship laptop, four smartphones, four IP cameras, and a LoRa transmitter. Under $5,000. You get object detection, face recognition, vehicle fingerprinting, mesh networking, ATAK integration, and AI threat assessment. No cloud, no vendor lock-in, and you can field it in weeks.
The ATO binder is three releases old. Make the build write the evidence.
A program office opens a vendor's accreditation package and finds a document written six months ago about software that no longer exists. The slowest part of fielding defense software is not writing it. It is proving it. EdgeLance made the proof a build step, generated from the release itself.
The Sovereignty Stack
In 1956 the United States stopped a British and French military operation without firing a shot. It just squeezed the pound. Rented capability comes with conditions, and nations discover the terms at the worst possible moment. AI is the next capability on loan, and it ships with three off switches. This paper defines the four tiers a nation must own to field intelligence that cannot be revoked, jammed, or metered.
Why the next C2 acquisition will not be a platform. It will be a layer.
Defense primes keep losing C2 deals because they bid platforms against platforms. The winning move is to own the tactical edge layer that sits underneath any platform. Acquiring that layer is faster than building it.
AirBridge: autonomous drone-assisted mesh for tactical data reachback (R&D concept)
R&D concept paper. A drone is more than an ISR asset. At the right altitude, it becomes a temporary airborne mesh bridge, store-and-forward courier, and reachback handoff point. This paper works through the range math, operational value, and routing model behind the AirBridge concept.
Top-down platforms give command a God view. Operators get a surveillance feed pointed at themselves.
Lattice and Maven are enterprise platforms built for command oversight and total data capture. The operator is not the customer; the operator is the data source. EdgeLance inverts that: mission intelligence belongs to the team, syncs upward by choice, and wipes clean when the op is over.
The plan left the building as a PDF
A planning cell spends twelve days synchronizing a plan on a live shared surface. Then the team steps off and the plan travels as a PDF, and at 0447 the first dead radio takes the synchronization with it. Planning software got staffs from days to hours. EdgeLance is the other half of that loop, the plan as a live mission workspace and the mission record flowing back as ground truth for the next plan.
Designing for the operator who does not trust you
SOF teams reject enterprise tools because every platform they have been handed is built for the people watching them, not for them. Designing for distrust means local-first by default, team-owned data, promotion-based sharing, and a system that proves it can be wiped before anyone trusts it.
The cost of battlefield AI: why localized compute on consumer hardware is the only model that scales
Cloud AI inference costs are rising 30-40% per model generation. Dedicated GPU hardware for the edge costs $15-25K per node with 6-month procurement timelines. Meanwhile laptop-class silicon ships 38 TOPS at $1,099 and phone-class chips match it in a pocket. The military that figures out how to run localized consumer AI compute in a coordinated mesh will own the next fight.
DARPA's MOSAIC concept needs a node-level operating layer. Nobody has built it yet.
MOSAIC warfare replaces monolithic platforms with modular nodes composed via AI networks. DARPA's 2026 RFI calls for autonomous drone constellations with edge-based computing and multi-agent operations. The concept is clear, but the node-level software layer that makes it work for ISR at the company level does not exist yet.
The locker becomes a fleet the moment every device knows what it can carry
An S6 shop stares at forty-three mixed devices bought over three years and a mission brief at 0900. Nobody knows what each one can actually run. EdgeLance enrollment fingerprints every device, scores it into a capacity tier, and sizes the mission pack to the node. The guesswork ends before the planning starts.
Chinese EW in the South China Sea already broke your cloud AI architecture
Six paved antenna sites at Mischief Reef. Five vehicle-mounted jammers at Subi Reef. GPS denial across four bodies of water. Russian EW in Ukraine cut precision weapon effectiveness by 90%. If your AI sends queries to the cloud, the adversary can disable it by attacking the link.
Tactical device management: turning 200,000 consumer devices into managed mission nodes
CMMC 2.0 requires 110 security controls across 200,000+ defense industrial base companies. NIST SP 800-124r2 separates MDM from mobile threat defense. No integrated solution handles multi-classification enforcement on consumer hardware with tactical features like RF suppression and duress wipe. This paper examines the gap and what it takes to close it.
The defense AI market is moving toward the tactical edge
Program offices want AI that works closer to the sensor and the operator, with less dependence on perfect connectivity. That opens a gap for platforms built around local inference, mesh routing, and managed COTS hardware.
The instructor cannot hear four nets at once. Coach can.
An instructor with a clipboard cannot grade four radio nets at the same time, so most comms errors die unrecorded. EdgeLance Coach transcribes live traffic on-device, attributes every transmission by voice fingerprint, and scores each call against doctrine with the audio anchored to every finding. Radio discipline finally has a scored target.
The ISR gap below the enterprise layer
Enterprise C2 and ISR platforms serve large programs well. The gap is lower in the formation, where teams need AI, mesh, cameras, and device control but have no dedicated infrastructure or program office backing them up.
AI model governance at the tactical edge: provenance, loadouts, and the auditability gap
DoD adopted five AI ethical principles in 2020. The RAI Toolkit mandates explainability for high-risk decisions. Then the Pentagon banned Anthropic from Maven and forced Palantir to rip out its core AI engine in 180 days. The gap between governance policy and operational tooling is where missions stall and legal reviews fail. This paper proposes a field-deployable model governance architecture.
What Ukraine taught NATO about consumer hardware in combat
Recent conflicts proved how fast commercial devices, drones, and compute show up in the field. COTS does not replace every military system. It needs a security and management layer before it belongs in a mission.
Two operators tasked the same drone. The software let them.
At shift change the bird belonged to whoever spoke last. Mission software fails when everything lives in one flat pile with no concept of ownership. EdgeLance binds every task, asset, and record to one of three scopes, and control leases make every handoff explicit. One owner, always, on the ledger.
The Operator Is the Platform
The most capable computer ever carried into combat did not arrive through a program of record. It arrived in the operator's pocket. A founding doctrine paper on what happens when the lowest unit of capability becomes a person, and what acquisition, training, and command must do about it.
AI API costs are about to explode and defense budgets are not ready
Cloud inference pricing is volatile and model demand keeps rising. Defense programs that rely on metered APIs inherit that uncertainty. Local inference puts the cost curve back in your hands.
Bottom-up ISR: why the next generation of tactical intelligence starts at the squad
DARPA's MOSAIC warfare concept, Ukraine's decentralized innovation model, and the FY26 budget all point in the same direction: pushing AI, sensors, and decision support lower in the force structure. This paper examines what ISR at the company level and below looks like without enterprise infrastructure.
Mesh networking for the dismounted warfighter: what works and what does not
Dedicated tactical radios still matter. EdgeLance Mesh sits above the transport, using whatever links are available and routing mission data by priority, bandwidth, battery, and trust.
A commander asked why the AI said that. The gateway is the answer.
Every model call in EdgeLance passes through one gateway that enforces capacity, classification and thermal policy before the model runs, then stamps a signed provenance record on the answer. Fourteen call sites, one door, and a receipt for every inference.
Local AI keeps working when links get contested
Contested links make cloud-only AI fragile. Local inference on edge hardware is how operators keep useful AI running when connectivity degrades.
Edge AI in contested spectrum: operational requirements for inference under EW denial
Chinese EW installations across the Spratly Islands, Russian GPS denial reducing precision weapon effectiveness by up to 90%, and Iranian autonomous drone production scaling 10x. This paper analyzes how electromagnetic threats drive the requirement for local inference and what that means for system architecture.
Why COTS hardware is becoming a serious tactical node
Modern consumer hardware has enough local compute for real edge AI workflows. The remaining problem is software: security posture, fleet management, model loadouts, and making operations mission-aware.
There is no cloud to cut
Russian jamming dropped a precision shell's hit rate from 70 percent to under 10. The link was the weakness, and every C2 system built in the last decade has the same one. EdgeLance keeps the mission in a signed ledger that every node holds, so jamming can shrink the mesh but never decapitate it.
Why Knox is not enough: classification-aware MDM for contested environments
Enterprise MDM proved consumer hardware can be managed. Tactical operations need more: data boundaries, emissions controls, NVG-compatible UI, duress workflows, and airgapped updates.
Acquisition pathways for edge AI platforms: OTAs, SWP, and the Barrier Removal Board
The March 2025 Hegseth software acquisition memo, the November 2025 acquisition transformation strategy, and DIU's 500+ OTA track record have reshaped how DoD buys software. This paper maps the acquisition pathways available to nontraditional defense AI vendors.
Why AI without source evidence is operationally useless
An AI that says 'hostile contact' without showing the camera clip, RF signature, and detection confidence is asking the operator to act on faith. That is not how tactical decisions work.
Trust is an architecture, not a policy
In January 1961 a hydrogen bomb fell on North Carolina and ran nearly its full arming sequence on the way down. One switch held. The nuclear enterprise answered by moving trust out of documents and into hardware, and every institution that handles dangerous capability has since made the same move. AI governance has yet to make it.
Mission continuity in contested comms: why every node has to be the system
Most edge platforms treat disconnected operation as a fallback. EdgeLance treats it as the baseline. Each node stays useful whether the network is degraded, intermittent, or denied entirely.
The network is already walking around
A dismounted team carries a dozen radios among them. Phones, watches, tablets, a laptop, maybe a $30 LoRa board in a cargo pocket. Every one of those devices can transmit. And none of them form a network, because no software treats them as one fabric. The hardware problem is solved and paid for. The software problem is wide open.
The defense stack has a hole in the middle. Soldiers fill it with group chats.
Defense software is consolidating into platforms for headquarters and pipes for the field. The space in between, where a team actually plans, runs and reviews a mission on its own hardware, has no serious product in it. The evidence is that soldiers keep building it themselves. A first attempt at sketching what belongs there.
Mission data should be mortal
Every mission system ever fielded treats accumulation as a virtue. Operators respond by keeping the real information out of the system. The fix is data that is born inside a mission and provably destroyed at the end.
Who owns the intelligence?
Militaries have always understood that some knowledge must die rather than be taken. Enterprise software forgot this. The case for mission-owned intelligence, provable deletion, and the AAR as the two ends of one lifecycle.
Renting versus owning intelligence
A military capability you rent is a capability someone else can turn off. AI is quietly becoming the most rented capability in defense, and there are three hands on the switch.
The cost of destruction collapsed. The software has not noticed.
A $1,000 drone killed a $250 million bomber in June. The most capable computer on the battlefield is already in every operator's pocket, and nobody has written the mission software for it.