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).
Module I/O — twelve modules
New modules (M, X, P) highlighted in green/teal. Each card: runs · input · output · key design note.
System-level gaps — updated
COMSEC resolved by Module X. Three showstoppers remain open.
Orchestration / Message Bus
No broker or execution graph defined across 12 modules.
Drone ↔ Server Comms Bandwidth
LoRa-class links cannot carry detection streams at scale. X adds ECDH overhead.
Data Contracts
No protobuf/JSON schemas at module boundaries — 12 modules now produce outputs.
M · Sensor Array Geometry
Multilateration needs ≥3 sensor nodes at known positions. Deployment constraint not captured.
Time Sync
TDOA (Module M) requires sub-microsecond time alignment across sensor nodes.
Latency Budget
No per-module ms allocation. 12 modules must fit inside G's <3s e2e target.
COMSEC → Module X · AES-256-GCM
AES-256-GCM authenticated encryption on drone→server mesh. ECDH key exchange per link per mission.
P Cold-Start (no T history)
P reads T history for PoL baseline — none exists at deployment day 1.
Final build — 15 June 2026. v2 (9 modules) + M + X + P = 12 modules. 12 → 19 models. 6 → 7 layers. COMSEC gap: addressed.
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.
Drone edge (Jetson)COCO/VisDrone pretrain → fine-tune military drone classes. Augment: low-light, motion-blur, thermal. TensorRT export mandatory for edge FPS.
Drone edge (Jetson)No learned weights — Kalman + IoU association. Tune thresholds for drone vibration. Output compressed for narrow mesh link.
DISC server (GPU)Same pretrain as A₁, medium/large variant. Fine-tune on LORROS thermal, long-range blur. Augment: haze, distance blur, FOV crop.
DISC server (GPU)Kalman + IoU. Stabilised LORROS input produces fewer lost tracks than on-drone path. No learned weights.
Ground station (CPU/GPU)ResNet-18/MobileNetV3. SNR augment (AWGN + Rician). Temperature scaling → calibrated confidence. Low-conf routes to LSTM branch.
Ground station (CPU)LSTM autoencoder on normal ambient RF (unsupervised). MSE reconstruction + rolling z-score. Zero-label cold-start.
DISC server (CPU)Classical TDOA mathematics. Hyperbola intersections + least-squares fit. Calibrate sensor array positions once at deployment. Degrades gracefully on node dropout.
DISC server (CPU)Delaunay triangulation + robust least-squares. Combines TDOA hyperbola intersections with AoA bearings for stability. No training data required.
Drone edge + DISC (both ends)AES-256-GCM: confidentiality + integrity in one pass. Hardware crypto on Jetson where available. GCM nonce prevents replay attacks.
Drone edge + DISC (both ends)Ephemeral ECDH (Curve25519). Keys rotated per-mission or on timer. Pre-load mission keys before deployment for disconnected ops.
DISC server (GPU)Coord transform pipeline: radar(rng/az) → camera(px) → VIO(local) → WGS84. Rule-based Mahalanobis gating. Learnable later with sufficient calibration data.
DISC server (GPU)Graph net on synthetic + labelled track pairs. Learned deduplication — not brittle threshold matching. Input format: geo-projected tracks (not BEV features).
DISC server (CPU)Deterministic FSM: tentative → confirmed → coasting → dropped. N-of-M confirmation (3/5). Kalman coasting through dropouts. Dual red/blue tables. History required by P.
DISC server (GPU)Siamese transformer, supervised bi-temporal pairs. Fine-tune on theatre imagery.
DISC server (GPU)Per-scene optimisation — not pretrained. Fits radiance field to captured views. Minutes to rebuild per mission, not real-time.
Forward drone (on-board)Tightly-coupled VIO + EKF for GPS blending. GPS-denied: VINS-Mono holds heading/position. Per-airframe calibration required.
DISC server (CPU)Gradient boosting. Bootstrap rules (drone & TTP<3min → high), operator-correction loop. Batch retraining only. P enrichment adds 2 new features.
DISC server (GPU)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.
DISC server (CPU)Unsupervised DBSCAN clustering + HMM state transitions. Self-builds "normal" baseline from T history. Flags statistically rare behaviour (z > 3.5 vs baseline) automatically.
DISC GPU + Convis voice pipelineSelf-hosted Sarvam-30B. Prompt-tuned for military reasoning + ROE check. Voice: Asterisk/FastAGI/Bulbul (Convis pipeline). RAG over ROE doc.
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.
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.
FRIENDLY 4
COVERAGE 78%
🔒 COMSEC OK
P: Checkpoint Alpha ⚠
Threat Priority · F · XGBoost+SHAP
P · Ambush Risk Overlay
3rd recon pass · z=4.1 vs baseline
Traj. Transformer → TTP 87s
G · Commander SITREP · Sarvam LLM
A₁ + E · Drone Management
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.
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
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.