AI Intelligence Layer · Restricted Access
1 · Architecture — Final Build

Input → Intelligence → Output Final

Nineteen models in twelve modules across seven layers. Three new modules close all remaining v2 brochure gaps: M · Multilateration (GPS-denied localization), X · COMSEC (AES-256-GCM encryption), P · Ambush & Pattern (behaviour prediction).

Psitech  ·  TC3 · DISC
Input · Sensors6
EO/IR · LORROS
Ground-based · stabilised · day+thermal
Drone Camera
On-board EO/IR · edge
New
Radar BFSR-SR
Sparse plots: range / az / Doppler
Verify
RF Detectors
300 MHz – 6 GHz IQ sweep
No HW
SAR / Satellite
RISAT-2B · bi-temporal · strategic
Drone IMU + GPS
Accel · gyro · GNSS → E + M
Blue Force — BFT / radio / manual → Track Manager (T).
Intelligence · DISC12 modules
Edge (on drone)
E · Positioning
VINS-Mono + GPS/EKF · single
A₁ · Drone Detection
YOLOv8-nano → ByteTrack · chain
Transport / Security
X · COMSEC New · closes mesh-encryption gap
AES-256-GCM · ECDH key mgmt · wrap
New
Detection
A₂ · Ground Detection
YOLOv8-m/l → ByteTrack · chain (server)
B · RF Signal Intelligence
CNN Spectrogram + LSTM Anomaly · merge
M · Multilateration New · ToA + Delaunay geo-fix
TDOA Engine + Delaunay Solver · localize
New
Fusion / State
C · Multi-Sensor Fusion
Late fusion → GNN Track Assoc. · chain
T · Track Manager
Lifecycle FSM · red + blue tables · stateful
Terrain (planning)
D₁ · Change Detection
ChangeFormer · standalone
D₂ · 3D Reconstruction
Gaussian Splatting · standalone
Decision
F · Threat Scoring
XGBoost + SHAP · fan-in · scoring only
P · Ambush & Pattern New · reads T history
Trajectory Transformer + PoL Miner · predict
New
⚠ Human Approval Gate
AI does perception & analysis — humans make decisions & actions. Nothing flows to actuation without operator approval.
Interface
G · Commander Interface
Sarvam LLM + Convis voice · reads all
Output · Commander7+2+1
Live Tactical
Common Operating Picture
2D map · red + blue ← T, E
Threat Alert Panel
Ranked · SHAP ← F, B
Ambush Risk Overlay
Predicted zones ← P
New
System Health / COMSEC
Link + encryption status ← X
New
Event Timeline
Detections, alerts, actions
Drone Management
Asset state ← A₁, E
RF Signal Monitor
Protocol · anomaly ← B
Planning Mode
3D Terrain View
Digital twin ← D₂
Change Overlay
Staging, construction ← D₁
Persistent Overlay
Voice / Text Interface
SITREP, COA ← G · any view

Module I/O — twelve modules

New modules (M, X, P) highlighted in green/teal. Each card: runs · input · output · key design note.

Edge
E · Positioning
VINS-Mono + GPS/EKF
single
RunsDrone (Jetson)
InputCamera + IMU + GPS
Output6-DOF pose → T, COP, Drone Mgmt, C
Edge
A₁ · Drone Det.
YOLOv8-nano → ByteTrack
chain
RunsDrone (Jetson)
InputDrone camera frames
OutputTrack IDs → X (COMSEC) → C
Transport · NEW
X · COMSEC
AES-256-GCM · ECDH
wrap
RunsDrone edge + DISC (both ends)
InputPlaintext A₁ + E output
OutputAES-GCM ciphertext → mesh → decrypt → C
Closes gap: "Zero encryption on mesh" (v2 gap card). Runs on both Jetson and DISC. ECDH per-link session keys, rotation per mission.
Detection
A₂ · Ground Det.
YOLOv8-m/l → ByteTrack
chain
RunsDISC GPU
InputLORROS RTSP stream
OutputGround track IDs → C
Detection
B · RF SIGINT
CNN + LSTM Anomaly
merge
RunsGround station
InputIQ sweep 300 MHz–6 GHz
OutputProtocol class + anomaly flag → C, Threat Panel
Detection · NEW
M · Multilateration
TDOA + Delaunay
localize
RunsDISC CPU
InputTimestamped arrivals from ≥3 RF/radar sensors
OutputWGS84 emitter fix (meter-level) → C as geo track
Closes gap: Brochure "ToA + Delaunay" was missing from v2. Not ML — classical TDOA geometry. Needs ≥3 sensors at known positions.
Fusion
C · Multi-Sensor
Late fusion → GNN
chain
RunsDISC GPU
InputA₁ (via X), A₂, radar, B, M geo-fix
OutputUnified tracks → T
Coord transform critical: radar(rng/az) → camera(px) → VIO(local) → WGS84. GNN deduplicates multi-sensor views.
State
T · Track Manager
Lifecycle FSM
stateful
RunsDISC CPU
InputC + E + Blue Force
OutputRed + blue table → COP, F, P, G, AAR
FSM: tentative → confirmed → coasting → dropped. Kalman coasting. Single source of truth — P reads history from T.
Terrain
D₁ · ChangeFormer
Change detection
standalone
RunsDISC GPU
InputBefore/after SAR pair
OutputChange masks → F (staging flag), Change Overlay
Terrain
D₂ · Gauss. Splatting
3D reconstruction
standalone
RunsDISC GPU
InputMulti-view drone imagery
Output3D twin → P (chokepoints), 3D Terrain View
Decision
F · Threat Scoring
XGBoost + SHAP
fan-in
RunsDISC CPU
InputT table (10+ features) + D₁ flag + P ambush risk
OutputPriority 1–5 + SHAP reason → Threat Panel, Gate
Decision · NEW
P · Ambush & Pattern
Traj. Transformer + PoL
predict
RunsDISC GPU + CPU
InputT track histories + D₂ terrain chokepoints
OutputPredicted paths + ambush zones → F enrichment, G, COP
Closes gap: Brochure "Ambush & Pattern Detection" now has architectural home. PoL Miner (DBSCAN + HMM) builds movement baseline from T history; Trajectory Transformer predicts intent.
Interface
G · Commander
Sarvam + Convis voice
reads all
RunsDISC GPU + voice pipeline
InputAll modules (event-driven + 60s heartbeat)
OutputSITREP, COA → voice overlay (<3s e2e)

System-level gaps — updated

COMSEC resolved by Module X. Three showstoppers remain open.

Showstopper

Orchestration / Message Bus

No broker or execution graph defined across 12 modules.

→ Kafka/NATS, topic schema, DAG scheduler
Showstopper

Drone ↔ Server Comms Bandwidth

LoRa-class links cannot carry detection streams at scale. X adds ECDH overhead.

→ Link budget spec, format, fallback
Showstopper

Data Contracts

No protobuf/JSON schemas at module boundaries — 12 modules now produce outputs.

→ Protobuf per interface, versioning
Required

M · Sensor Array Geometry

Multilateration needs ≥3 sensor nodes at known positions. Deployment constraint not captured.

→ Survey, calibration procedure
Required

Time Sync

TDOA (Module M) requires sub-microsecond time alignment across sensor nodes.

→ PTP/GPS-NTP, max skew budget
Required

Latency Budget

No per-module ms allocation. 12 modules must fit inside G's <3s e2e target.

→ Allocate per-module, measure
Addressed

COMSEC  → Module X · AES-256-GCM

AES-256-GCM authenticated encryption on drone→server mesh. ECDH key exchange per link per mission.

✓ Architecture home: Module X (Transport / Security layer)
Required

P Cold-Start (no T history)

P reads T history for PoL baseline — none exists at deployment day 1.

→ 72h observe-only mode, synthetic priors from D₂

Final build — 15 June 2026. v2 (9 modules) + M + X + P = 12 modules. 12 → 19 models. 6 → 7 layers. COMSEC gap: addressed.

2 · Model Reference — Final

Every model: input → output → training Final

Filter by module. New modules M, X, P highlighted. INPUT in blue, OUTPUT in green. Training strategy and data source per card.

Showing 19 models across 12 modules
Edge
A₁ · Drone Detection
YOLOv8-nano
Per-frame detector — on drone
RunsDrone edge (Jetson)
InputDrone on-board camera frame (each)
OutputBoxes + class + confidence → ByteTrack
Training

COCO/VisDrone pretrain → fine-tune military drone classes. Augment: low-light, motion-blur, thermal. TensorRT export mandatory for edge FPS.

DataVisDrone, DOTA + AirSim synthetic renders.
Zone 1edge
Edge
A₁ · Drone Detection
ByteTrack
Multi-object tracker — edge
RunsDrone edge (Jetson)
InputYOLOv8-nano detections per frame
OutputStable track IDs → X · COMSEC (compressed)
Design

No learned weights — Kalman + IoU association. Tune thresholds for drone vibration. Output compressed for narrow mesh link.

DataMOT17 validation. No training data.
Zone 1edge
New
A₂ · Ground Detection
YOLOv8-m/l
Per-frame detector — LORROS, server
RunsDISC server (GPU)
InputLORROS stabilised video (RTSP stream)
OutputBoxes + class + confidence → ByteTrack (A₂)
Training

Same pretrain as A₁, medium/large variant. Fine-tune on LORROS thermal, long-range blur. Augment: haze, distance blur, FOV crop.

DataVisDrone, DOTA + LORROS calibration footage (must collect).
Zone 1server
New
A₂ · Ground Detection
ByteTrack (server)
Multi-object tracker — LORROS
RunsDISC server (GPU)
InputYOLOv8-m/l detections (stabilised feed)
OutputGround track IDs → C
Design

Kalman + IoU. Stabilised LORROS input produces fewer lost tracks than on-drone path. No learned weights.

DataMOT17 validated. No training data.
Zone 1server
B · RF Signal Intelligence
CNN Spectrogram
RF protocol classifier
RunsGround station (CPU/GPU)
InputSTFT spectrogram 128×128 of 100–200 ms IQ capture
OutputProtocol class (OcuSync/Analog FPV/FHSS/Unknown) + confidence → C, Threat Panel
Training

ResNet-18/MobileNetV3. SNR augment (AWGN + Rician). Temperature scaling → calibrated confidence. Low-conf routes to LSTM branch.

DataDroneRF (Tulsa), DeepSig RML2016, GNU Radio synthetic.
Zone 1server
B · RF Signal Intelligence
LSTM Anomaly
Wideband novelty sentinel
RunsGround station (CPU)
Input300 MHz–6 GHz sweep → 512 freq-bins × 20 time steps
OutputAnomaly flag + offending freq bins (z > 3.5) → C, Threat Panel
Training

LSTM autoencoder on normal ambient RF (unsupervised). MSE reconstruction + rolling z-score. Zero-label cold-start.

DataSelf-collected ambient recording (2–3 days). No labels needed.
Zone 1server
New · M
M · Multilateration
TDOA Engine
Time-difference-of-arrival localization
RunsDISC server (CPU)
InputTimestamped signal arrivals from ≥3 RF sensor / radar nodes (nanosecond precision)
OutputHyperbolic intersection → candidate emitter position (lat/lon/alt estimate)
Design (not ML)

Classical TDOA mathematics. Hyperbola intersections + least-squares fit. Calibrate sensor array positions once at deployment. Degrades gracefully on node dropout.

DataNo training data. Sensor positions + calibration capture. Requires ≥3 overlapping coverage zones.
Zone 1server
New. Closes brochure "ToA + Delaunay" feature. Not ML — pure geometry.
New · M
M · Multilateration
Delaunay Solver
Triangulation stability & geo-fix
RunsDISC server (CPU)
InputTDOA candidate positions + AoA bearing lines from RF detectors
OutputWGS84 position fix (meter-level accuracy) → C as geo-located emitter track
Design (not ML)

Delaunay triangulation + robust least-squares. Combines TDOA hyperbola intersections with AoA bearings for stability. No training data required.

DataGeometry + calibration only. Deploy RF array with ≥3 known-position nodes.
Zone 1server
New · X
X · COMSEC
AES-256-GCM
Authenticated encryption — mesh link
RunsDrone edge + DISC (both ends)
InputPlaintext detections + telemetry from A₁ and E on drone
OutputCiphertext + GCM auth tag → mesh → verify + decrypt at DISC → C
Design (not ML)

AES-256-GCM: confidentiality + integrity in one pass. Hardware crypto on Jetson where available. GCM nonce prevents replay attacks.

DataNo training data. Key material + nonce management only.
Zone 1edge + server
Closes gap: "Zero encryption on mesh" (v2 gap). Transport-level, not application-level.
New · X
X · COMSEC
ECDH Key Mgmt
Per-link session key exchange
RunsDrone edge + DISC (both ends)
InputNode identity certificates / pre-shared roots
OutputPer-link session keys + rotation schedule + replay protection → AES-256-GCM
Design (not ML)

Ephemeral ECDH (Curve25519). Keys rotated per-mission or on timer. Pre-load mission keys before deployment for disconnected ops.

DataPKI / pre-shared roots. No training data. Hardware secure enclave on Jetson for key storage.
Zone 1edge + server
Reworked
C · Multi-Sensor Fusion
Late Fusion Engine
Track-level association — joins 5 inputs
RunsDISC server (GPU)
InputA₁ tracks (via X), A₂ tracks, radar plots, B RF class, M geo-fix
OutputMulti-sensor candidate tracks in common WGS84 geo-frame
Strategy

Coord transform pipeline: radar(rng/az) → camera(px) → VIO(local) → WGS84. Rule-based Mahalanobis gating. Learnable later with sufficient calibration data.

DataCalibrated multi-sensor captures (must collect). Start rule-based.
Zone 2server
Updated
C · Multi-Sensor Fusion
GNN Track Assoc.
Cross-sensor deduplication
RunsDISC server (GPU)
InputCandidate tracks from Late Fusion Engine
OutputOne unified track per real-world target → T
Training

Graph net on synthetic + labelled track pairs. Learned deduplication — not brittle threshold matching. Input format: geo-projected tracks (not BEV features).

DataSynthetic graphs + labelled overlapping-coverage scenes.
Zone 2server
v2-New
T · Track Manager
Track Lifecycle FSM
Authoritative track table — single source of truth
RunsDISC server (CPU)
InputUnified tracks from C + 6-DOF pose from E + Blue Force positions
OutputRed + blue track table → COP, F, P, G, Timeline, AAR persistence
Design (not ML)

Deterministic FSM: tentative → confirmed → coasting → dropped. N-of-M confirmation (3/5). Kalman coasting through dropouts. Dual red/blue tables. History required by P.

DataConfig + threshold tuning only. No training data. No labels.
Zone 2server
Split
D₁ · Change Detection
ChangeFormer
Temporal change — strategic layer
RunsDISC server (GPU)
InputBefore/after SAR or satellite image pair
OutputChange masks (tire tracks, construction, staging) → F (staging flag), Change Overlay
Training

Siamese transformer, supervised bi-temporal pairs. Fine-tune on theatre imagery.

DataLEVIR-CD, OSCD, S2Looking + custom AO pairs.
Zone 1server
Split
D₂ · 3D Reconstruction
Gaussian Splatting
3D terrain — planning mode
RunsDISC server (GPU)
InputMulti-view drone imagery (per-mission capture)
Output3D digital twin → P (terrain chokepoints), 3D Terrain View (planning)
Design

Per-scene optimisation — not pretrained. Fits radiance field to captured views. Minutes to rebuild per mission, not real-time.

DataNo dataset — optimises per mission capture.
Zone 1server
Moved
E · Positioning
VINS-Mono + GPS/EKF
Visual-inertial odometry with GPS fusion
RunsForward drone (on-board)
InputCamera + IMU + GPS (PX4/ArduPilot)
Output6-DOF pose over mesh → T, C, COP, Drone Mgmt
Training

Tightly-coupled VIO + EKF for GPS blending. GPS-denied: VINS-Mono holds heading/position. Per-airframe calibration required.

DataEuRoC/TUM-VI validation. Per-airframe calibration capture.
Zone 1edge
Narrowed
F · Threat Scoring
XGBoost + SHAP
Interpretable scoring — scoring only
RunsDISC server (CPU)
InputT track table (10+ features: position, velocity, class conf, closure rate, RF flag, M geo-fix, P ambush enrichment, D₁ staging flag)
OutputPriority 1–5 + SHAP reason per contact → Threat Alert Panel, Gate cue
Training

Gradient boosting. Bootstrap rules (drone & TTP<3min → high), operator-correction loop. Batch retraining only. P enrichment adds 2 new features.

DataDISC logs + rule labels + operator feedback over time.
Zone 2server
New · P
P · Ambush & Pattern
Trajectory Transformer
Intent forecasting — next-position prediction
RunsDISC server (GPU)
InputTrack histories from T (red position sequences) + terrain chokepoints from D₂
OutputPredicted paths + ambush-risk zones → F enrichment, G context, Ambush Overlay on COP
Training

Self-supervised next-position prediction over track sequences (LLM-style over trajectory tokens). Fine-tune on labelled ambush/infiltration events. Augment with D₂ terrain chokepoints as prior.

DataArchived T mission tracks + labelled ambush events. Bootstrap: synthetic chokepoint scenarios + D₂ terrain priors.
Zone 1server
New. Closes brochure "Ambush & Pattern Detection". Reads T history — cold-start: 72h observation-only mode.
New · P
P · Ambush & Pattern
PoL Miner
Pattern-of-life baseline & deviation detection
RunsDISC server (CPU)
InputAggregated T track history per grid cell / time-of-day
OutputNormal-movement baseline + deviation flags (staging, massing, repeated recon) → F enrichment, G
Training

Unsupervised DBSCAN clustering + HMM state transitions. Self-builds "normal" baseline from T history. Flags statistically rare behaviour (z > 3.5 vs baseline) automatically.

DataSelf-built from T history persistence. No labels needed for baseline. Improves continuously in production.
Zone 1server
Updated
G · Commander Interface
Sarvam LLM
Natural-language commander — persistent voice overlay
RunsDISC GPU + Convis voice pipeline
InputAll modules event-driven + 60s heartbeat: T state, F ranking, P prediction, B RF class, D terrain, X COMSEC status
OutputSITREP + COA + voice answers → Voice Overlay (<3s e2e target)
Training

Self-hosted Sarvam-30B. Prompt-tuned for military reasoning + ROE check. Voice: Asterisk/FastAGI/Bulbul (Convis pipeline). RAG over ROE doc.

DataFew-shot + RAG. No fine-tune to start. Synthetic SITREP/COA pairs for alignment.
Zone 1server
3 · Data-flow — Live

How data moves through all 12 modules

v1-style continuous simulation — dots flow on every edge simultaneously. New nodes: M (Multilateration), X (COMSEC), T (Track Manager), P (Ambush & Pattern). Node glows when data arrives.

camera / detection radar / fused RF / SIGINT SAR / terrain drone IMU / VIO threat / geo-fix COMSEC / encrypted track state
4 · Output

What the commander sees

One screen — not forty. The 2D COP (Track Manager) is the primary view. Threat Panel (F+SHAP), Ambush Overlay (P), COMSEC status (X), SITREP voice overlay (G), and Drone Management (A₁+E) are persistent panels.

LIVEDISC · SECTOR 4--:--:--🔒 X · AES-256-GCM OKMIL-STD-2525
CONTACTS 0
FRIENDLY 4
COVERAGE 78%
🔒 COMSEC OK
P: Checkpoint Alpha ⚠

Threat Priority · F · XGBoost+SHAP

P · Ambush Risk Overlay

Checkpoint Alpha · MEDIUM RISK
3rd recon pass · z=4.1 vs baseline
Traj. Transformer → TTP 87s

G · Commander SITREP · Sarvam LLM

Awaiting decision — nothing actuates until approved.

A₁ + E · Drone Management

5 · Risk Register

Bottlenecks & how each is solved

Ordered by severity. Each risk is paired with a concrete architectural decision already in the final build. Updated with M, X, P-specific risks.

6 · Cold-start

When data is scarce — get it or generate it

Military data is classified and sparse. Six of twelve modules need no training data at all (M, X, T, LSTM, Gaussian Splatting, PoL Miner). Build order: zero-label first.

Strategy — cheapest first

  • Zero-label modules first (6 of 12). M (geometry), X (crypto), T (FSM), LSTM-B (unsupervised), D₂ (per-scene), P-PoL (DBSCAN baseline) — all runnable in week one without a single label.
  • Bootstrap from rules. XGBoost (F) starts on heuristic labels (drone & TTP<3min → high), then operator corrections refine it in production. P-Transformer bootstraps from D₂ terrain priors.
  • Public datasets to pretrain. DroneRF, DeepSig, VisDrone, MOT17, nuScenes, LEVIR-CD cover feature extractors — fine-tune only last layers on a small theatre set.
  • Synthetic generation. GNU Radio for RF protocols; AirSim for EO/IR; synthetic ToA arrivals for TDOA calibration; terrain chokepoint scenarios for P cold-start.
  • Track Manager unlocks the data flywheel. T's history persistence is what eventually feeds P retraining, F retraining, and AAR review. Build T early.

Per-module data

Build order: F (rule labels, no sensors) → B-LSTM (ambient recording) → B-CNN (public datasets) → A₁/A₂ (VisDrone fine-tune) → M (deploy array, calibrate) → C (calibrated capture) → T (needs C) → P (needs T history, 72h+) → X (PKI setup before field deployment).
7 · Use Case — Real World

Hostile FPV Cross-Sector Infiltration · Sector 4

All 12 modules responding to one contact event. Each step shows which modules are Parallel (independent, no waiting), Sequential (depends on predecessor), or Join (waits for all inputs before proceeding).

Scenario

Northern Sector 4, early morning. A hostile FPV quadcopter (OcuSync3 control link) crosses the ridge at 46m AGL heading southeast toward Forward Post 2. Four drones airborne, LORROS on north mast, BFSR-SR and RF array deployed at three known grid positions.