Project WildGuard
Protecting remote critical infrastructure, whether guarding a perimeter, securing electrical substations, or monitoring endangered wildlife reserves, requires continuous surveillance in environments with zero cloud connectivity and strict solar/battery power budgets.
Project WildGuard is a fully autonomous, off-grid ambient intelligence platform. It processes up to 8 live streams of camera feeds simultaneously using edge-native vision models on the Metis hardware.
Instead of straining weak satellite or radio networks by attemting to stream heavy video frames, an integrated local Small Language Model translates visual threat detection data into compressed, conversational text summaries.
A ranger or tactical officer miles away can query the system over a low-bandwidth LoRA mesh network via text, receiving instant, intelligent situational updates entirely offline.
Technical Architecture
This project maximizes the Metis PCIe card inference power, keeping the power draw envelope constrained for efficient consumption.
- The Vision Layer: Utilizing the Voyager SDK to process parallel video streams running an INT8-quantized YOLO object detection model, tracking tactical and ecological classes: e.g [unauthorized_vehicle, poacher, crawling_person, smoke_plume, weapon, wildlife_event, restricted_activity].
- The Embedding & Analytics Engine (New): Instead of simple text logging, visual events are parsed into structured semantic metadata and passed through a lightweight, edge-optimized embedding model. These vectors are stored in a local, offline vector database directly in the system's memory. This enables semantic anomaly detection (flagging patterns that deviate from historical baseline vectors) and trend analytics without cloud overhead.
- The Cognitive Layer (Local RAG): When a user submits a query, the system generates an embedding of the question, performs a fast vector similarity search against the local database to retrieve relevant historical context from the captured data, and feeds that exact context into the SLM context window.
- The LoRA Bridge: Text summaries are compressed and broadcasted via a connected low-power radio transceiver, eliminating cellular or internet dependencies. This also has the potential to leverage low-power satellital modems for critical alerts, updates and command execution from remote control stations.
Conversational Analytics
Adding embeddings unlocks complex analytical queries that standard vision pipelines simply cannot answer:
-
Trend Analysis Query: "Compare vehicle activity in Sector 4 between this week and last week. Is there a spike?"
-
System Action: The system uses the local vector DB to aggregate and count semantic matches for "unauthorized vehicle" in Sector 4 across the historical timeline, calculating the percentage delta entirely locally.
-
Output: "Vehicle activity in Sector 4 has increased by 35% compared to last week. Most breaches are occurring between 02:00 AM and 04:00 AM. I have flagged a recurring white pickup truck vector."
-
-
Semantic / Contextual Search: "Show me any suspicious activity around the perimeter fencing during the rainstorm yesterday."
-
System Action: Even if the log doesn't explicitly contain the word "suspicious," the embedding model matches the semantic intent to entries like "crawling person detected near boundary wire" during the weather timestamp.
-
Output: "At 16:45 PM yesterday during heavy rainfall, Camera 2 registered a high-probability perimeter breach vector (crawling posture). The target retreated after 14 seconds."
-
Voyager Wingman
The complexity of stitching asynchronous, parallel vision pipelines to a token-generation language model loop is reduced by leveraging Voyager Wingman to generate the optimal structural YAML configurations, manage memory layout between the streamed inputs and the axllm context window, and accelerate debugging so I can build a stable deployment in record time.
The Industrial Scaling Strategy
While this project focuses on rugged eco-conservation and perimeter defense, the underlying technology serves as a direct blueprint for commercial industry. A perimeter breach at a remote asset uses the exact same vision-to-text pipeline required to protect a commercial manufacturing yard, an automotive factory fence line, or a sprawling solar farm. Mastering this multi-modal edge integration provides an immediate pathway to pivot into high-value industrial asset monitoring.
Execution Plan and Deliverables
During this month sprint, I will deliver a complete, reproducible project containing:
- Weeks 1-2: Build the parallel YOLO vision streams and metadata logging structure using Voyager Wingman Prompts.
- Week 3: Compile the SLM using the axllm toolkit and integrate the local text-query pipeline.
- Week 4: Build the local Web UI dashboard, test the text queries and data transmission on the LoRA network, and finalize documentation.
Final Deliverables: A fully open-source GitHub repository containing all source code, deployment configurations, the Wingman prompt journey logs and a video demo showcasing the system successfully running real-time queries entirely offline.
