Aaryan Patwardhan

SentinelMesh — INSPIRON 5.0 (CSI COEP)

• —Designed MAPE-K autonomic loop for self-healing autonomous drone fleet
• —YOLOv8n at 55fps on RTX 3050 Ti with dual CUDA co-inference (vision + LLM)
• —Built Adversarial Debate Engine: Agent-A Dispatcher vs Agent-B Skeptic before any dispatch
• —Implemented confidence tier gating — LLM bypassed at conf ≥ 0.88 for ~50ms dispatch

Ghost-Admin — Autonomous Server Healing Daemon

• —Built 7-stage MAPE-K closed-loop: trend-based pre-detection, behavioral fingerprinting, multi-signal context fusion, semantic intent classification, graceful degradation ladder, pre-kill forensic dumps, and SIEM-ready JSONL audit log
• —llama3.2:3b via Ollama (CUDA) classifies process intent across 5 categories including UNDER_ATTACK — giving the daemon rudimentary intrusion detection built in
• —RAG memory layer (FAISS + sentence-transformers) retrieves similar past incidents before every AI query — daemon improves with each event
• —Reduced mean time to detection from ~15 minutes (on-call) to under 30 seconds; zero cloud API calls

Aura — Passive Attendance Intelligence

• —Ingested RADIUS Accounting logs from campus WLC — no student app, no QR codes, no cooperation required
• —Primary session key is 802.1X-authenticated User-Name, not MAC address — immune to MAC randomization (iOS 14+, Android 10+)
• —Isolation Forest Focus Score detects bandwidth anomaly patterns multivariant on (bytes, duration) jointly — no hardcoded thresholds per course type
• —Full stack: FastAPI ingestion (async, <100ms), Redis session state, PostgreSQL persistence, React dashboard with role-based views

Upwork Automation Pipeline

• —Built two-version automated job-hunting pipeline with LLM scoring and Telegram alerts
• —V2: composite scoring (50% AI relevance, 25% client quality, 25% competition opportunity)