I build at the intersection of edge AI, industrial computer vision, cybersecurity, and full-stack systems. Currently deploying production-grade AI on NVIDIA Jetson AGX Orin and DGX Spark — from factory-floor defect detection and PLC-integrated automation to security LLMs fine-tuned on threat intelligence. I also harden mobile apps and build encrypted communication tools.

• OXmaint Edge Portal — Production industrial AI on Jetson Orin: 12 YOLO models, PLC (Allen-Bradley, Arduino Opta), ROS2 robot simulation
• WAZUHLLM — Fine-tuned security LLM with Wazuh SIEM + CTI pipeline (NVD/KEV/ThreatFox) + RAG vault for SOC automation
• Industrial CV models on NVIDIA TensorRT — conveyor defects, spill, dust, PPE, galvanizing, thermal fault detection
• Multi-tenant CCTV AI — Real-time surveillance with automated work order creation and escalation pipelines
• Encrypted messaging PWA with QR-based E2E key exchange and WebRTC calling
• Voice-to-text automation for hands-free computing on Linux
• One-command Ubuntu workstation setup — trackpad gestures, systemd services, the works

Fine-tuning LLMs on threat intelligence to build an AI-native SOC analyst.

CTI Feeds (NVD / KEV / ThreatFox / blog sources)
        │
        ▼
   Ingestion Pipeline ──── ChromaDB RAG Vault
        │                        │
        ▼                        ▼
  Qwen 3.6 35B + LangGraph ──► Structured Threat Reports
        │
        ▼
  Wazuh SIEM Integration ──► Automated Alert Triage

• LLM: Qwen 3.6 35B via Ollama, fine-tuned with unsloth on threat corpus
• RAG: LangGraph agent with ChromaDB-backed retrieval over NVD CVEs, KEV catalog, ThreatFox IOCs
• SIEM integration: Wazuh alert enrichment — LLM triages alerts, suggests response actions, generates hunt hypotheses
• Knowledge base: Khoj + SilverBullet for SOC playbook management and retrieval
• Architecture: security-llm-architecture.md — full system design documented

Edge AI Platform — Architected and deployed a real-time computer vision platform on NVIDIA Jetson Orin running 12 YOLO models in production. Integrated PLCs (Allen-Bradley, Arduino Opta) and ROS2 robot simulation into a unified factory monitoring system serving multiple industrial customers, enabling defect detection at conveyor speeds with sub-second inference latency.

AI-Powered SOC Automation — Built a fine-tuned security LLM pipeline (Qwen 3.6 35B + LangGraph + ChromaDB RAG) that ingests NVD, KEV, and ThreatFox threat intelligence to automate SIEM alert triage, reducing mean-time-to-respond for security incidents.

Cross-Platform Deployment — Production workloads span NVIDIA Jetson AGX Orin (edge) and DGX Spark GB10 (data center), with models optimized via TensorRT and served behind Cloudflare zero-trust tunnels. On-device LLM inference with Ollama/vLLM for factory-floor decision support without cloud dependency.

Multi-Tenant AI Surveillance — Designed and deployed a CCTV AI system doing real-time hazard detection with automated work order creation and escalation pipelines, currently monitoring multiple industrial sites.

Systems Automation — Created a one-command Ubuntu workstation setup used by migrating engineers, and voice-to-text tooling that eliminates manual transcription for hands-free factory-floor documentation.

• Security LLM fine-tuning — unsloth + Qwen on CTI feeds for automated SOC triage and hunt hypothesis generation
• Adversarial robustness in industrial CV models — can a sticker on a conveyor belt fool a defect detector?
• On-device LLM inference (vLLM, Qwen, Gemma 4) for real-time factory safety decision support across Orin + DGX Spark
• Secure inference pipelines — running CV models without exposing sensitive factory floor data
• Multi-Spark clustering — InfiniBand/MPI orchestration across DGX Spark fleet for distributed LLM inference
• NVIDIA DeepStream 9.0 — multi-stream video analytics with 3D pose tracking and anomaly detection