Applied AI & Robotics Engineer. Developing ROS 2 autonomous systems, autonomous LLM agents, and PyTorch vision pipelines

Kunal Somani  • Peerlist

Autonomous mine inspection rover built without touching physical hardware. Full digital twin first approach using NVIDIA Isaac Sim. 1. PPO-trained locomotion in PyBullet with 200K training timesteps 2. ROS 2 sensor nodes for IMU, LiDAR, and BLDC encoders 3. Full 3D visual simulation in NVIDIA Isaac Sim 5.1 with live sensor feeds. 160m tunnel traversal at 3.00 m/s, 95% complete, fully autonomously 4. SLAM via slam_toolbox and RTAB-Map for real-time underground mapping 5. Crack detection, gas sensing, and spark-proof chassis design Next milestone: Bridging Isaac Sim to ROS 2 for live RViz 2D SLAM map generation during exploration

Canary Rover, Autonomous Mine Inspection Syst

Production-deployable recursive autonomous agent built on the OODA loop for automated web task completion. 1. Playwright handles JS-heavy DOMs for real browser interaction 2. Claude Tool Use synthesizes a Python solution just-in-time per task 3. RestrictedPython and SIGALRM sandbox execution for safe code running 4. HTTP submission loops until terminal state is reached. Fully autonomous. 5. Durable async jobs via ARQ on Redis with retry logic 6. Prometheus, Loki, and Grafana monitor latency and throughput across worker containers

Helix, Recursive Autonomous Web Agent

Production-deployable instruction-to-deployment backend. Takes a natural language prompt and pushes a live website to GitHub Pages autonomously. 1. Hybrid RAG pipeline combining Cohere dense retrieval with BM25 sparse search, fused via Reciprocal Rank Fusion for context retrieval 2. Anthropic Tool Use generates schema-validated code via structured output constraints 3. GitHub App deployer pushes live sites using short-lived installation tokens. Zero manual deployment steps. 4. FastAPI and Celery orchestrate async task execution. Redis Pub/Sub streams live logs over WebSocket to a React and TypeScript dashboard. 5. Full observability via Prometheus, Grafana, and OpenTelemetry across the async worker pipeline

Archon, Autonomous Code Generation and Deploy

Production-deployable fully local AI assistant with zero cloud dependency. 1. Runs Whisper STT, moondream2 VQA, Kokoro TTS, and Phi-3 LLM entirely on local hardware 2. BART-MNLI zero-shot semantic router dispatches queries across three processing paths: personal knowledge retrieval, OS-level tool execution, and general conversation 3. Hybrid RAG using MiniLM dense retrieval via Qdrant plus BM25 sparse, reranked by a cross-encoder 4. GBNF grammar-constrained sampling reduces JSON parse failures and tool hallucination to near zero 5. Next.js frontend with FastAPI orchestrator, fully containerized with Docker

Axon Core, Self-Hosted Tri-Modal AI Assistant

Classical computer vision pipeline for cross-verifying digital RFID FASTag claims against physical vehicle geometry at toll plazas. 1. 3780-dimensional HOG feature vectors extracted per vehicle image 2. LinearSVC trained on 24K+ images achieving 97% accuracy on multi-axle vehicle classification 3. Cross-modal centroid tracker with MOG2 virtual tripwire for lane-level vehicle detection 4. Misclassification error analysis with a simulated audit pipeline 5. Streamlit operator dashboard for real-time monitoring