// intelligence layer
Summit.OS ships with eleven production-ready ONNX models custom-built and trained by BigMT.ai — domain-expert feature engineering, purpose-built training pipelines, and real-world public datasets blended with scenario-specific synthetic data. They run entirely on your hardware — no cloud, no API calls, no data leaves your network.
// flagship model
Real-time object detection from drone cameras, CCTV feeds, and satellite imagery. YOLOv8-architecture ONNX model with 18 mission-critical detection classes.
YOLOv8n-based · 640×640 input · 18 detection classes · <15ms inference on CPU
// full model inventory
All eleven models were designed, trained, and exported by BigMT.ai — not fine-tuned from
third-party checkpoints. Training scripts live in packages/ml/
and are fully open. Every model ships as ONNX — no PyTorch, no GPU required at runtime.
Given a detected event (class + coordinates + risk level), predicts the optimal mission type to dispatch. Feeds the tasking engine to auto-generate missions from sensor detections. Supports progressive fine-tuning from real operator-approved mission history.
Assigns a threat level to any detected event based on class, confidence, location context, and environmental factors. Drives alert severity and PENDING_APPROVAL gating for high-stakes missions.
Binary classifier that suppresses noisy detections before they reach the operator alert queue. Trained on historical confirmed vs. discarded alert patterns. Reduces operator alert fatigue.
Predicts whether an active alert will escalate in severity within the next observation window. Used to pre-position assets and alert commanders before a situation worsens.
Groups spatially and temporally related alerts into unified incidents. Prevents operators from being flooded with duplicate alerts for the same underlying event (e.g., fire spread across multiple sensor zones).
Estimates the probability of mission success before launch. Considers asset capabilities, environmental conditions, mission type, and target characteristics. Informs human-in-the-loop DISPATCH decisions.
Optimal asset-to-mission matching. Given available assets (drones, ground units, sensors) and pending missions, scores each pairing and surfaces the best allocation. Accounts for proximity, fuel, payload, and capability.
LSTM-based model that detects anomalous entity behavior in time-series tracks. Flags unusual movement patterns, loitering, erratic flight paths, or unexpected sensor dropouts that rule-based filters miss.
Assesses operational risk from weather conditions: wind speed, precipitation, visibility, temperature, and lightning proximity. Used to auto-hold missions when conditions exceed safe operating thresholds.
Generates 128-dimensional visual embeddings for detected objects. Enables cross-camera entity re-identification — track a person or vehicle as they move between non-overlapping camera fields of view.
YOLOv8-based visual detector with 18 mission-critical classes. See full class list above. Primary inference engine for video feeds from drones, CCTV, and aerial imagery. Runs at real-time frame rates on CPU.
// where the data came from
BigMT.ai built custom training pipelines and domain-expert synthetic data generators covering 10+ operational domains — wildfire, search & rescue, maritime, infrastructure, agriculture, hazmat, and more. Public real-world datasets were downloaded, preprocessed, and blended with scenario-specific synthetic samples to cover edge cases and low-frequency events. No proprietary data, no user data, no scraped content.
Global wildfire and fire front detections updated every 10 minutes from MODIS and VIIRS satellite instruments. Primary source for smoke/fire/fire_front training samples.
1.5M+ severe weather records including floods, tornadoes, hurricanes, and hail. Primary source for weather risk and escalation predictor training.
Pipeline failure and hazardous materials incident reports from 2010–present. Source for pipeline_damage, oil_spill, and chemical_plume detection labels.
Global Biodiversity Information Facility — 2B+ wildlife occurrence observations. Used for dangerous_animal and wildlife-related detection class training.
Real-time global earthquake feed. Used for structural risk scoring and incident correlation training data for ground movement events.
Earth Observatory Natural Event Tracker — volcanoes, floods, wildfires, landslides, dust storms, and sea/lake ice events with georeferenced coordinates.
Crowd-sourced wildlife observations including threatened and endangered species. Supplementary source for ecological detection classes and rare event distributions.
Global air quality monitoring data used as a proxy for fire smoke concentration and chemical plume dispersion during training.
Global Disaster Alert and Coordination System — earthquakes, cyclones, floods, volcanoes, and tsunamis with severity scoring. Used for risk scorer calibration.
UN OCHA humanitarian reporting database — floods, earthquakes, epidemics, conflicts. Used for high-level incident type classification training.
Low-frequency scenarios (e.g., simultaneous multi-hazard events, extreme sensor noise) generated via generate_data.py to ensure model robustness at the margins.
// optional external network calls
Summit.OS makes no outbound calls by default. The following are opt-in integrations, all disabled unless you set the corresponding env var. None of these calls send your operational data — they only pull in public reference feeds.
| Service | What it does | Enabled by | What's sent |
|---|---|---|---|
| OpenSky OPTIONAL |
Live ADS-B aircraft positions for the spatial intelligence overlay | OPENSKY_ENABLED=true |
Your bounding box coordinates (or nothing for worldwide). No entity data sent outbound. |
| CelesTrak OPTIONAL |
TLE orbital elements for satellite position propagation on the 3D globe | CELESTRAK_ENABLED=true |
HTTP GET request for TLE data file only. No user data sent. |
| MaxMind OPTIONAL |
GeoIP database auto-update for geoblocking enforcement | MAXMIND_LICENSE_KEY= set |
Authenticated download of GeoLite2-Country.mmdb. No user IP addresses sent. |
| YubiCloud OPTIONAL |
OTP validation for YubiKey hardware authentication (Yubico OTP protocol) | YUBICO_CLIENT_ID= set |
YubiKey OTP token sent to Yubico's validation servers. No user data. Can be replaced with self-hosted YK-VAL server. |
| Infisical / Vault OPTIONAL |
Secrets management backend for production deployments | INFISICAL_TOKEN= or VAULT_ADDR= set |
Secret key names only — no operational data, no user data. |
| S3 OPTIONAL |
Audit log shipping for SOC2 compliance retention | AWS_AUDIT_BUCKET= set |
Append-only audit log entries (operator actions, auth events). Only your own bucket — Anthropic/BigMT never receives this data. |
// bring your own data
All training scripts are open source in packages/ml/.
Fine-tune any model on your own operator history for higher accuracy in your specific environment.
Every alert confirm/discard and mission outcome is logged to the audit trail in NDJSON format. Export your deployment's history or label new samples using your own sensor data.
Run the public data downloader to augment your dataset with fresh global event data.
cd packages/ml && python download_real_data.py
Each model has its own training script. Example for the risk scorer:
python train_risk_scorer.py
Visual detector and re-ID embedder training are Colab notebooks (GPU recommended):
packages/ml/colab/colab_visual_detector.ipynb
packages/ml/colab/colab_reid_embedder.ipynb
Replace the corresponding .onnx file in packages/ml/models/ and restart the intelligence service.
No code changes required — all services load models from the path in PLANNER_MODEL_PATH or the default models directory.
docker compose restart intelligence
# Run all tabular model training scripts in sequence for script in train_risk_scorer train_mission_classifier train_false_positive_filter \ train_escalation_predictor train_incident_correlator train_outcome_predictor \ train_asset_assignment train_sequence_anomaly train_weather_risk_scorer; do echo "=== Training $script ===" python packages/ml/${script}.py done