SAR Proof of Concepts (POC)

AI-Powered COSPAS-SARSAT Distress Beacon Registration Voice Agent

Architected and engineered a completely offline, local-first interactive voice response (IVR) control loop to automate the data intake, validation, and storage of emergency distress beacons (EPIRBs, ELTs, PLBs) for injection into a National Beacon Registry database. Designed to address life-safety tracking data liabilities for emergency agencies and commercial fleets, the agent bypasses fragile keyword matchers by treating voice data collection as a dynamic relational matrix. The platform pairs a rigid, context-aware state machine with a local LLM to support natural, compound verbal error corrections without dropping the active call session or risking database hallucinations.

Metrics at a Glance:

Key Architectural Achievements:

• Context-Aware State Machine Workflows: Implemented an adaptive, dynamic sequential slot-filling configuration loop that evaluates prerequisites on the fly, automatically skipping irrelevant infrastructure categories (such as bypassing marine questions for an aviation ELT beacon) to dramatically cut call times.
• Natural Multi-Intent Correction Routing: Engineered an advanced natural language routing engine that leverages a lightweight local LLM to decode unstructured conversational user corrections mid-readback (e.g., "Wait, my phone number and email are both wrong"), intelligently mapping them back to correct backend database keys.
• Anti-Hallucination Schema Guardrails: Enforced strict operational boundaries by limiting the LLM to an analytical structural sorting role; used Python list comprehensions to filter model outputs against active session candidate keys, guaranteeing the AI can never directly alter or fabricate system database values.
• High-Performance Offline Voice Processing Core: Integrated a zero-dependency, ultra-low-latency local pipeline pairing high-speed neural transcriptions with fast-inference text-to-speech engine blocks, ensuring total compliance with rigid institutional data privacy standards.
• Deterministic Data Storage Ledger: Developed a reliable state checkout system that automatically assigns unique 8-character verification tracking tokens, appends explicit ISO-8601 timestamps, and commits transaction payload arrays securely into a structured local ledger file.

Intelligent MEOLUT Ground Segment Telemetry Calibration & Real-Time AI Pipeline

Designed and engineered an enterprise-grade, end-to-end processing pipeline that unifies traditional satellite signal processing with deep learning and Generative AI for Medium Earth Orbit Local User Terminals (MEOLUT). The system ingests authentic ASCII Two-Line Element (TLE) satellite catalogs, computes 3D Keplerian trajectory vectors, and isolates ionospheric tracking anomalies via an unsupervised PyTorch Autoencoder. A live OpenCV operator dashboard maps real-time coordinate streams, visually flags data link degradation, and triggers a LangChain orchestration layer to automate executive-level root-cause engineering diagnostics.

Metrics at a Glance:

Key Architectural Achievements:

• Authentic Orbital Feature Engineering: Programmed a parsing module using NumPy to digest raw satellite TLE tracking blocks, computing geometric Slant Ranges and true Doppler FOA frequency adjustments relative to a ground station ECEF frame.
• Unsupervised Deep Learning Calibration: Engineered an unsupervised PyTorch Autoencoder to model nominal tracking baselines and dynamically capture signal degradation without relying on brittle, hardcoded threshold limits.
• Reactive Operator Visualization Scope: Constructed a real-time monitor canvas using OpenCV that mirrors polar radar grid arrays, dynamically traces tracking lines, and flips UI states into an alarm mode when processing faults arise.
• Deterministic GenAI Remediation Routing: Architected an automated diagnostic layer using LangChain Expression Language (LCEL) that converts complex mathematical errors into structured, actionable root-cause engineering tickets.
• Decoupled Production Architecture: Designed the full application workspace into a clean package layout, separating core computation nodes from the rendering engine to enable future Apache Kafka and RAG scaling hooks.

CoSAR-TS: Coalition Search & Rescue Multi-Agent Proof of Concept

Engineered a runnable, deterministic reconstruction of a classical multi-agent system (MAS) architecture inspired by the foundational AAAI-04 research paper. The platform bypasses modern LLM reliance to demonstrate the auditability and power of symbolic AI, handling capability-based resource matchmaking, declarative policy enforcement, and multi-incident asset contention. Built on top of a decoupled message-passing bus architecture, the engine processes real-world constraints such as airspace restrictions, calculates triage-weighted priorities, and coordinates complex logistics across fictitious allied territories into a unified operator dashboard.

Metrics at a Glance:

Key Architectural Achievements:

• Capability-Based Semantic Matchmaking: Programmed an independent profile matching module that scans available resource capabilities against dynamic incident requirements, autonomously ranking assets based on operational readiness and raw ETA metrics.
• Declarative Policy Enforcement Engine: Built a rule engine inspired by KAoS that evaluates standalone boundary and cross-border airspace constraints, proving system intelligence by vetoing structurally optimal assets when policy-denied.
• Multi-Casualty Contention Arbitration: Architected decentralized client entities utilizing the Contract Net Protocol to issue bids for identical, scarce coalition helicopters alongside an independent resolver agent resolving asset locks via triage priority.
• Human-in-the-Loop Operational View: Constructed a multi-view Streamlit control room dashboard mapping live agent interactions, geometric incident coordinates, spatial asset trajectories, and comprehensive reasonings for human escalations.
• Decoupled Asynchronous Infrastructure: Designed a clean Python workspace dividing core computation components from UI code via a central message bus to guarantee zero shared state and facilitate future RAG planning extensions.

Agentic RAG COSPAS-SARSAT Search & Rescue System (Mission Control Center POC)

Architected and engineered an open-source, local multi-agent geospatial intelligence prototype designed to automate distress beacon telemetry parsing and triage compliance based on COSPAS-SARSAT standards. Built a decoupled, hard-boundary ingestion data pipeline that extracts ground-truth metrics from 15-Hex and 23-Hex beacon strings using deterministic bitwise arithmetic, passing it as structured JSON to autonomous agents. The platform unifies the local Llama 3.2 model with a vectorized database of official operating handbooks, establishing a secure, air-gapped system that streams live execution logs into a responsive interface while generating mission-ready alert templates.

Metrics at a Glance:

Key Architectural Achievements:

• Hard-Boundary Decoupling Ingestion: Prevented transcription vulnerabilities and LLM hallucinations by extracting binary tracking configurations (MMSI, TAC, Serial) using deterministic bit-level parsing instead of open text inference.
• C/S T.018 Structural Verification: Configured rigorous bitwise safety loops that validate bit-1 status, check 12–14 bit frame synchronization patterns (`101`), and flag alerts automatically if Type Approval Certificates drop below 10,000.
• International MID Country Mapping: Integrated an automated dictionary look-up array that instantly translates numerical Maritime Identification Digits (e.g., `503`) into clear country text strings (`Australia`) to prevent routing delays.
• ConcurrentNode Queue Isolation: Resolved background multi-threaded state-management collisions within the web framework by building a vanilla Python queue pipeline that streams agent traces safely to the primary UI rendering thread.
• Failsafe Traceability Framework: Engineered an explicit verification boundary that injects visible warning banners (`[UNVERIFIED - NO HANDBOOK REFERENCE AVAILABLE]`) directly into final briefs if local manuals are unpopulated or unreached.