KAVACH: An On-Device Tactical Intelligence Platform with SLM360-Powered Natural Language C2, Real-Time ISR, and Automated Reporting

1. Introduction

1.1 The Tactical AI Gap

Modern military operations generate massive volumes of multi-source intelligence data -sensor feeds, reconnaissance imagery, unit positions, threat reports -that exceed human cognitive bandwidth for real-time synthesis. Existing Command and Control (C2) systems address this through cloud-connected platforms requiring persistent satellite or network links. In forward-deployed, denied, or degraded communications (DDIL) environments, these systems become inoperable, leaving operators with manual processes, paper maps, and voice radio.

The gap is acute at the tactical edge: squad-to-platoon level units that need real-time situational awareness, threat assessment, and decision support but lack dedicated intelligence infrastructure. These units operate with:

1. No network connectivity -forward positions, patrol bases, observation posts
2. Limited computing hardware -smartphones and tablets, not servers
3. Time-critical decisions -seconds matter for threat response and maneuver
4. Multi-source data -visual feeds, unit reports, map data, threat intelligence
5. Small team size -no dedicated intelligence analyst at squad level

1.2 Limitations of Existing Tactical Systems

Current tactical C2 platforms suffer from fundamental constraints:

System	Connectivity	Hardware	AI/ML	Cost
ATAK/TAK	Requires mesh/satellite	Android (no AI)	None	Government-only
CPOF/C2PC	Requires TOC server	Desktop workstation	Limited	$500K+ per node
Palantir TITAN	Cloud-connected	Server cluster	Cloud ML	Enterprise pricing
Commercial GIS	Internet required	Laptop/tablet	None	$10K-100K/year

None of these systems provide on-device AI inference for real-time ISR processing, automated reporting, or intelligent route planning without network connectivity.

1.3 Our Contribution

We present KAVACH, a tactical intelligence platform that runs entirely on a commercial Android device with zero network dependency. KAVACH integrates seven AI-powered modules into a unified tactical operating picture:

• Tactical C2 with natural language command interface powered by SLM360
• Data Fusion with multi-source intelligence knowledge graph
• ISR Processing with real-time object detection from camera/drone feed
• Auto SITREP with AI-generated SALUTE reports from battlefield data
• Patrol Optimization with threat-aware route planning
• Threat Intel with predictive threat analysis and anomaly detection
• SLM360 Engine providing the core NLU backbone (577K parameter transformer)

All AI models total 848KB and run entirely on-device using the SLM360 NanoEncoder + BaseDecoder architecture, achieving real-time inference with zero network calls.

---

2. System Architecture

2.1 Overview

KAVACH follows a modular architecture where each tactical function operates as an independent module sharing a common AI backbone (SLM360) and data layer (tactical knowledge graph).

KAVACH System Architecture

+------------------+------------------+------------------+------------------+
|   TACTICAL C2    |   DATA FUSION    |  ISR PROCESSING  |   AUTO SITREP    |
|  Force display   | Knowledge graph  | Object detection |  SALUTE reports  |
|  NL commands     | Entity linking   | Camera/drone     |  Threat reports  |
+------------------+------------------+------------------+------------------+
|  PATROL OPTIM.   |   THREAT INTEL   |  SLM360 ENGINE   |                  |
|  Route planning  | Anomaly detect   | 577K transformer |                  |
|  Threat avoid    | Pattern analysis | NanoEncoder+Dec  |                  |
+------------------+------------------+------------------+------------------+
|                    TACTICAL DATA LAYER                                    |
|            Units / Threats / Terrain / Routes / Reports                   |
+--------------------------------------------------------------------------+
|                    SLM360 AI BACKBONE (848KB)                             |
|         NanoEncoder (384-dim) + BaseDecoder (25 intents)                  |
+--------------------------------------------------------------------------+
|                    ANDROID PLATFORM (API 26+)                             |
|              TFLite + CameraX + MapView + Sensors                        |
+--------------------------------------------------------------------------+

2.2 Design Principles

KAVACH is built around four core principles:

1. Zero network dependency: Every computation runs on-device. No cloud calls, no mesh network required, no satellite link needed. The system is fully operational in airplane mode.
2. Sub-second response: All AI inference completes within the operator's decision cycle. NLU intent classification in <50ms, ISR detection at >15 FPS, report generation in <10 seconds.
3. Military-grade UX: Tactical color scheme (dark background, green/red symbology), NATO military symbols, grid coordinate system, and SALUTE-format reporting.
4. Minimal footprint: Total AI model size of 848KB enables deployment on any mid-range Android device without storage or memory constraints.

2.3 Platform and Dependencies

Component	Details
Target platform	Android 8.0+ (API 26), optimized for landscape tablet
ML runtime	TensorFlow Lite 2.14.0 with 4-thread CPU inference
NLU engine	SLM360 NanoEncoder + BaseDecoder (577K parameters)
Vision model	YOLOv8n TFLite (12MB, float16 quantized)
Map renderer	Custom tactical overlay on satellite/terrain tiles
Total model size	848KB (NLU) + 12MB (vision)
Camera framework	AndroidX CameraX 1.4.1
Total APK size	~45 MB

---

3. SLM360 NLU Engine

3.1 Architecture

The core AI backbone of KAVACH is the SLM360 NanoEncoder + BaseDecoder, a lightweight transformer-based NLU engine optimized for tactical command understanding.

Model specifications:

Parameter	Value
Architecture	NanoEncoder (6-layer transformer) + BaseDecoder (2-layer classifier)
Total parameters	577K
Model size	848KB (INT8 quantized)
Embedding dimensions	384
Vocabulary	Tactical domain (500 tokens)
Intents	25 tactical command classes
Inference time	<50ms on mobile CPU

3.2 Tactical Intent Taxonomy

KAVACH's NLU engine recognizes 25 tactical intents organized into 6 categories:

Category	Intents	Examples
Situational Awareness	show_threats, show_units, overview, unit_status	"Show me all threats", "Unit status"
Reporting	sitrep, threat_report, patrol_report, contact_report	"Generate SITREP", "Threat report"
Fire Support	call_for_fire, fire_mission, adjust_fire, cease_fire	"Call for fire on grid 6200-7352"
Maneuver	move_unit, set_waypoint, route_plan, rally_point	"Move Tiger to grid 6400-7100"
Assessment	threat_assessment, coa_analysis, terrain_analysis	"Threat assessment for sector bravo"
Control	mark_friendly, mark_hostile, set_boundary, set_phase_line	"Mark hostile at bearing 045"

3.3 Command Processing Pipeline

Operator Input (text/button)
    |
    v
SLM360 NanoEncoder (384-dim embedding)
    |
    v
BaseDecoder (25-class softmax)
    |
    v
Intent + Confidence Score
    |
    v
Entity Extraction (grid coords, unit names, threat types)
    |
    v
Command Execution (update map / generate report / plan route)

The NLU pipeline processes commands in two phases:

1. Intent classification: The NanoEncoder produces a 384-dimensional embedding of the input text, which the BaseDecoder classifies into one of 25 tactical intents with a confidence score.
2. Entity extraction: A pattern-based extractor identifies tactical entities (grid coordinates, unit callsigns, threat types, directions) from the classified command.

3.4 Quick-Action Buttons

For high-tempo operations where typing is impractical, KAVACH provides pre-mapped tactical quick-action buttons:

• Show Threats -display all known hostile contacts on map
• SITREP -generate situation report from current tactical picture
• Unit Status -display status of all friendly units
• Call for Fire -initiate fire support request workflow
• Show Units -display all friendly force positions
• Threat Assessment -generate composite threat analysis
• COA -course of action analysis
• Overview -full tactical overview of area of operations

Each button maps directly to an SLM360 intent, bypassing NLU processing for instant execution.

---

4. Tactical C2 Module

4.1 Map Display

The Tactical C2 module provides a real-time common operating picture with:

Force disposition:

• Friendly units (blue): displayed with NATO military symbols (rectangle for mechanized, diamond for reconnaissance, triangle for infantry, circle for HQ)
• Hostile contacts (red): displayed with inverted NATO symbols with threat type labels (SNIPER, PATROL, UAV, MORTAR, MECHAN)
• Unit labels: callsign names (Bravo, Recon, HQ, Tiger, Echo, Charlie, Alpha)

Tactical overlays:

• Sector boundaries (dashed lines dividing AO into quadrants)
• Phase lines and control measures
• Threat zones (shaded red areas for known hostile positions)
• Route overlays (patrol routes, supply routes)

4.2 Grid Coordinate System

All positions use military grid reference system (MGRS) format:

Grid Reference: 6200-7352
Format: EASTING-NORTHING (4-digit)
Precision: 100m grid squares

4.3 Natural Language Interface

The Tactical C2 module accepts both button-press and natural language commands processed by the SLM360 engine. On initialization, the system displays:

SYSTEM READY. Select a command below.
SLM360 NanoEncoder + BaseDecoder loaded successfully.
25 tactical intents active. All inference on-device.

Example interactions:

Command	Intent	Action
"Show Threats"	show_threats	Highlight all hostile contacts on map
"Unit Status"	unit_status	Display readiness of all friendly units
"Call for Fire"	call_for_fire	Open fire support request form
"Threat Assessment"	threat_assessment	Generate composite threat analysis

---

5. ISR Processing Module

5.1 Architecture

The ISR (Intelligence, Surveillance, and Reconnaissance) Processing module provides real-time object detection from the device camera or connected drone feed using a custom-trained YOLOv8n model.

Model specifications:

Parameter	Value
Architecture	YOLOv8n (nano)
Input resolution	320 x 320 x 3 (RGB, float32)
Output format	[1, 8, 2100] -2100 predictions x 8 values
Classes	PERSONNEL, VEHICLE (mapped to tactical categories)
Model size	12 MB (TFLite, float16 quantized)
Confidence threshold	0.35
NMS IoU threshold	0.45

5.2 Detection Pipeline

The ISR module processes camera frames in real-time:

Camera/Drone Feed (30 FPS)
    |
    v
YUV -> Bitmap Conversion
    |
    v
Resize to 320x320
    |
    v
YOLOv8n TFLite Inference (4 threads)
    |
    v
NMS + Confidence Filtering
    |
    v
Tactical Overlay Rendering
    |
    v
Detection Log Update

5.3 Tactical Display

The ISR display provides:

• Live camera feed with detection bounding boxes (green brackets for PERSONNEL)
• Confidence scores displayed alongside each detection (e.g., "PERSONNEL 63%")
• Detection log with timestamped entries showing class, confidence, and bounding box coordinates
• Statistics bar: total targets, personnel count, vehicle count, and FPS counter
• LIVE indicator confirming real-time processing status

Example detection log output:

DETECTION LOG
20:31:45  PERSONNEL
  conf: 66%  box:[0.75,0.36,1.00,1.00]
20:31:44  PERSONNEL
  conf: 67%  box:[0.72,0.41,1.00,1.00]
20:31:43  PERSONNEL
  conf: 62%  box:[0.73,0.36,1.00,1.00]

TARGETS: 9 | PERSONNEL: 9 | VEHICLES: 0 | FPS: 19

5.4 Performance

Metric	Value
Inference speed	~19 FPS on mid-range Android
Detection classes	PERSONNEL, VEHICLE
Confidence range	47-73% (observed)
Detection log	50-entry circular buffer with timestamps
Model size	12 MB

---

6. Auto SITREP Module

6.1 Report Generation

The Auto SITREP module generates standardized military reports from the current tactical picture using the SLM360 engine. Reports follow the SALUTE format used by NATO and Indian armed forces:

SALUTE Format:

Field	Description	Example
Size	Number and composition of observed force	2x patrol elements
Activity	What the observed force is doing	Route clearance and observation
Location	Grid coordinates of observation	Grid 6200-7352, Sector East/West
Unit	Identification of observed unit	Tiger Patrol, Recon Team
Time	Date-time group of observation	DTG 202031FEB26
Equipment	Equipment observed	Standard infantry kit, NV equipment

6.2 AI Assessment

Following the structured SALUTE fields, the module generates an AI assessment synthesizing the available intelligence:

AI ASSESSMENT:
  cobra team conducting clearance activity observed patrol
  grid 2ring north

The AI assessment is generated by the SLM360 decoder operating on the structured tactical data, producing a natural language summary of the situation.

6.3 Report Types

Three report types are available:

Report	Content	Use Case
GENERATE SITREP	Full situation report with all unit positions and threat status	Periodic situation updates to higher HQ
THREAT REPORT	Focused report on hostile contacts, threat levels, and recommended responses	Immediate threat notification
PATROL REPORT	Route-specific report with observations, terrain, and contact information	Post-patrol debriefing

6.4 Performance

Metric	Value
Report generation time	~8 seconds on-device
Report format	SALUTE (NATO standard)
AI assessment	Natural language synthesis via SLM360
Network requirement	None (fully offline)

---

7. Patrol Optimization Module

7.1 Route Planning

The Patrol Optimization module computes threat-aware patrol routes that avoid known hostile positions while maintaining coverage of the assigned sector.

Route computation:

1. Waypoint selection: Operator designates start/end points and mandatory waypoints
2. Threat overlay: Known hostile positions generate threat zones (circular exclusion areas)
3. Route optimization: A* pathfinding with threat-weighted cost function
4. Terrain analysis: Terrain type (open field, urban, forested) affects route scoring
5. Output: Optimal route with waypoints, ETA, risk assessment, and terrain description

7.2 Route Display

The route display provides:

• Sector grid divided into named quadrants (ALPHA, BRAVO, CHARLIE, DELTA)
• Friendly unit positions with callsign labels
• Hostile contacts with threat zone overlays (shaded red circles)
• Computed route as a connected polyline through numbered waypoints (W1-W6)
• Route details panel with risk level, ETA, waypoint count, terrain description, and threat avoidance information

Example route details:

ROUTE DETAILS
ROUTE ALPHA -Sector Patrol
Risk: LOW
ETA: 2h 15m
Waypoints: 6
Terrain: Mixed terrain: hills, sparse vegetation
Threat avoid: 1 zone

Mixed terrain: hills, sparse vegetation. Good cover.

7.3 Threat Avoidance

The route planner incorporates threat avoidance through a weighted cost function:

cost(segment) = distance(A, B)
              + threat_weight * threat_proximity(segment)
              + terrain_weight * terrain_difficulty(segment)

Where:

• threat_proximity increases cost exponentially as the route approaches a known hostile position
• terrain_difficulty penalizes open terrain (higher exposure) and rewards covered terrain
• threat_weight and terrain_weight are operator-configurable parameters

---

8. Data Fusion Module

8.1 Knowledge Graph

The Data Fusion module maintains a tactical knowledge graph that links entities across all sensor modalities and intelligence sources:

Entity types:

Entity	Attributes	Sources
Friendly Unit	Callsign, position, status, strength, equipment	C2 module, manual input
Hostile Contact	Type, position, threat level, last seen, activity	ISR, patrols, intelligence
Terrain Feature	Type, position, cover rating, passability	Map data, patrol reports
Route	Waypoints, risk, ETA, terrain, last used	Patrol optimization
Report	Type, time, content, source unit	Auto SITREP

8.2 Entity Linking

The knowledge graph automatically links related entities:

• A hostile contact detected by ISR is linked to its corresponding threat zone in the patrol optimizer
• A patrol report referencing a grid coordinate is linked to all entities at that location
• Unit status changes propagate to route ETAs and threat assessments

---

9. Threat Intel Module

9.1 Threat Analysis

The Threat Intel module performs predictive threat analysis by:

1. Pattern detection: Identifying recurring hostile activity patterns (time of day, location clustering, movement corridors)
2. Anomaly detection: Flagging unusual activity that deviates from established patterns
3. Threat scoring: Computing composite threat levels from multiple intelligence sources
4. Prediction: Estimating likely hostile positions and actions based on historical patterns

9.2 Threat Levels

Level	Criteria	Response
LOW	Single unconfirmed report, distant location	Monitor
MEDIUM	Confirmed activity, moderate proximity	Increase awareness
HIGH	Multiple sources confirm, close proximity	Prepare response
CRITICAL	Imminent threat, weapons engagement range	Immediate action

---

10. SLM360 Engine Module

10.1 Live Inference Demo

The SLM360 Engine module provides a live demonstration of the NLU backbone, allowing operators to:

• Test natural language commands and observe intent classification in real-time
• View confidence scores and entity extraction results
• Benchmark inference latency on the current device
• Verify that all 25 tactical intents are loaded and active

10.2 Model Specifications

Parameter	Value
Architecture	NanoEncoder (6-layer) + BaseDecoder (2-layer)
Parameters	577K total
Quantization	INT8
Model file size	848KB
Embedding dimension	384
Tactical intents	25
Inference latency	<50ms (P50)

---

11. Results and Discussion

11.1 System-Level Performance

Metric	Value
Total APK size	~45 MB
Total AI model size (NLU)	848KB
Total AI model size (Vision)	12 MB
NLU inference	<50ms
ISR processing	~19 FPS
SITREP generation	~8 seconds
Route computation	<3 seconds
Cold start time	<5 seconds
Network requirement	None (100% offline)
Minimum hardware	Android 8.0, ARM64, 3GB RAM

11.2 NLU Performance

The SLM360 NanoEncoder + BaseDecoder achieves reliable intent classification across all 25 tactical commands:

Metric	Value
Intent accuracy	>95% on tactical command set
Entity extraction	Grid coordinates, callsigns, threat types
Inference time	<50ms on mobile CPU
Model size	848KB (INT8)
Active intents	25

11.3 ISR Performance

Metric	Value
Inference speed	~19 FPS on mid-range Android
Detection classes	PERSONNEL, VEHICLE
Confidence range	47-73% (observed in field testing)
Model size	12 MB (float16 TFLite)
Input resolution	320 x 320

11.4 Limitations

1. ISR detection range: At 320x320 input resolution, detection confidence drops significantly beyond ~100m for personnel-sized targets. Higher resolution (640x640) would improve range at the cost of FPS.

2. NLU vocabulary: The 25-intent taxonomy covers core tactical operations but does not include specialized functions like MEDEVAC requests, engineer tasks, or logistics operations. These are planned for future versions.

3. Auto SITREP quality: The SLM360 decoder produces concise assessments but with limited vocabulary and occasional grammatical imprecision. The structured SALUTE fields remain accurate as they are template-driven.

4. Map data dependency: While AI inference is fully offline, the tactical map requires pre-cached map tiles for the area of operations. Tile caching must be performed before deployment to DDIL environments.

5. Single-device operation: The current version operates on a single device. Multi-device tactical mesh networking for shared situational awareness is planned for future development.

---

12. Comparison with Existing Systems

Capability	KAVACH	ATAK	Palantir TITAN	CPOF
On-device AI	Yes (848KB)	No	No	No
ISR processing	Real-time (19 FPS)	None	Cloud-based	None
NLU commands	25 intents, <50ms	None	Cloud NLP	None
Auto reporting	SALUTE in ~8s	Manual only	Cloud-generated	Manual only
Route planning	Threat-aware, <3s	Basic waypoints	Cloud-optimized	Manual
Network required	None	Mesh/satellite	Always	Always
Hardware	Android phone/tablet	Android phone	Server cluster	Workstation
Total AI models	848KB + 12MB	N/A	GB-scale (cloud)	N/A
Cost	Commercial Android	Government-issued	Enterprise ($M)	Military ($500K+)

---

13. Future Work

1. Expanded intent taxonomy: Add 25+ additional intents covering MEDEVAC, logistics, engineer tasks, air support coordination, and NBC (nuclear/biological/chemical) reporting.
2. Multi-device mesh: Enable peer-to-peer tactical data sharing between KAVACH instances over Bluetooth or ad-hoc Wi-Fi for shared situational awareness without internet.
3. Voice commands: Integrate on-device speech recognition for hands-free operation during movement.
4. Enhanced ISR: Upgrade to YOLOv8s or YOLOv8m for improved detection range and add vehicle sub-classification (wheeled, tracked, technical).
5. Predictive routing: Use historical patrol data and threat patterns to pre-compute optimal routes for common patrol areas.
6. Integration with ATAK: Develop a plugin architecture to enable KAVACH AI modules to run within the ATAK ecosystem for units already equipped with TAK devices.

---

14. Conclusion

KAVACH demonstrates that a comprehensive tactical intelligence platform -encompassing command and control, ISR processing, automated reporting, threat analysis, and route planning -can operate entirely on a commercial Android device with zero network dependency.

By leveraging the SLM360 NanoEncoder + BaseDecoder architecture (577K parameters, 848KB), KAVACH provides natural language tactical command understanding in under 50ms. Combined with YOLOv8n-based ISR processing at ~19 FPS, SALUTE report generation in ~8 seconds, and threat-aware route planning in under 3 seconds, the system delivers decision-quality intelligence at the tactical edge.

KAVACH's fully on-device architecture makes it uniquely suited for:

• Forward-deployed units operating in DDIL environments
• Border patrol and surveillance operations without network infrastructure
• Rapid deployment -any soldier with a smartphone becomes a tactical intelligence node
• Training exercises where classified network access is not available

The 848KB NLU model size and 12MB vision model ensure minimal storage footprint, enabling deployment alongside existing military applications on standard-issue devices. Future work will focus on multi-device mesh networking, expanded intent coverage, and voice command integration for hands-free tactical operation.

---

References

• [1] U.S. Army. "ATP 6-0.5: Command Post Organization and Operations." Headquarters, Department of the Army, 2023.
• [2] NATO. "AJP-3.9: Allied Joint Doctrine for Joint Targeting." NATO Standardization Office, 2023.
• [3] Jocher, G., Chaurasia, A., & Qiu, J. "Ultralytics YOLOv8." Ultralytics, 2023.
• [4] TAK Product Center. "Android Team Awareness Kit (ATAK)." U.S. Air Force Research Laboratory, 2024.
• [5] Palantir Technologies. "Palantir TITAN: AI-Enabled Targeting and ISR." Palantir Defense, 2024.
• [6] Kumar, S. and Agrawal, P. "SLM360: A Lightweight Semantic NLU Engine Achieving State-of-the-Art Accuracy with Sub-50ms Latency." 360Labs, 2025.