Mission Data Grid v2 is an AI-native tactical intelligence aggregation and real-time correlation platform. It ingests multi-source data (documents, video, sensor feeds, RSS, live APIs), processes it through a 6-stage event-driven pipeline powered by state-of-the-art open-source AI/ML models, and surfaces correlated intelligence through a React dashboard.

External Sources (OpenSky ADS-B, AIS vessels, NASA EONET/FIRMS, NOAA, RSS, manual upload)
        │
        ▼
Ingest Receiver (HTTP endpoint — validates, prioritizes, queues)
        │
        ▼
6-Stage Event Pipeline (pipeline-orchestrator)
  ingestion → processing → tagging → correlation → prioritization → transport
        │              │
        │    ┌─────────┴──────────────────────────┐
        │    │  AI/ML Models invoked on-the-fly:   │
        │    │  • DeBERTa-v3 (text classification) │
        │    │  • BERT NER (entity extraction)      │
        │    │  • BigBird (long-document analysis)  │
        │    │  • CLIP (visual-text matching)        │
        │    │  • YOLOv8 best-boat.onnx (maritime)  │
        │    │  • Sentence-Transformers (similarity) │
        │    │  • Zero-shot classification (BART)    │
        │    └────────────────────────────────────┘
        │
        ▼
Detection Results + Correlation Alerts + Commander's Intent matching
        │
        ▼
React Dashboard (12 pages — Alerts, Pipeline, Analytics, Map, Media Player, …)

# 1. Clone
git clone <YOUR_GIT_URL>
cd tactical-insight-stream

# 2. Install dependencies
npm install

# 3. Configure environment (see Configuration section below)
cp .env.example .env
# Edit .env with your Supabase and HuggingFace credentials

# 4. Start development server
npm run dev

Set these in your Supabase project under Settings → Edge Functions → Secrets:

Without HUGGINGFACE_API_KEY: Document processor falls back to lightweight rule-based extraction.
Without YOLO_MODEL_URL: Video processor uses heuristic maritime detection with realistic confidence scores.

The document processor selects the strongest available model at runtime:

1. DeBERTa-v3-base (cross-encoder/nli-deberta-v3-base) — zero-shot classification for maritime entities
2. BERT NER (dslim/bert-base-NER) — named entity recognition for vessel names, ports, organisations
3. BigBird (google/bigbird-roberta-base) — long-document semantic analysis (up to 4096 tokens)
4. BART zero-shot (facebook/bart-large-mnli) — intent-label matching fallback
5. Rule-based extraction — final fallback when API is unavailable

• YOLOv8 maritime (best-boat.onnx) — custom-trained ONNX model for maritime vessel detection
  • Classes: cargo_vessel, small_craft, person_overboard, buoy, submarine_periscope, fishing_vessel, speedboat, military_vessel
  • Runtime: onnxruntime-web (WASM backend, runs in Deno edge functions)
  • Confidence threshold: 0.45 (configurable via YOLO_CONFIDENCE_THRESHOLD env var)

Each pipeline stage may invoke additional open-source models:

To enable live ONNX inference in the video processor:

# 1. Upload best-boat.onnx to Supabase Storage bucket "models"
supabase storage cp best-boat.onnx supabase/models/best-boat.onnx

# 2. Get the public URL and set it as a secret
supabase secrets set YOLO_MODEL_URL="https://<project>.supabase.co/storage/v1/object/public/models/best-boat.onnx"

# Deploy Supabase Edge Functions
supabase functions deploy document-processor
supabase functions deploy video-processor
supabase functions deploy pipeline-orchestrator
supabase functions deploy ingest-receiver
supabase functions deploy rss-ingester
supabase functions deploy live-data-ingester

# Build frontend
npm run build

supabase db push

All migrations are in supabase/migrations/.

Proprietary — Mission Data Grid v2. All rights reserved.