Two Pillars
Physical AI
Intelligence that senses, decides, and acts in the physical world — cognitive layers for industrial assets, stateful modeling, and real-time anomaly detection.
Edge AI
Intelligence that runs where the data is born — quantized TFLite models on microcontrollers, sub-100ms inference latency, and hybrid cloud-edge systems.
Digital Twins
Where the pillars meet: a live model of the physical system, fed by edge data and reasoned over in real time — state and behavior, not just metrics.
Pillar 1: Physical AI — Digital Twin Systems
Industrial Digital Twins run on a five-layer, event-driven stack that bridges low-latency field hardware with cloud-scale analytics. My hands-on production experience covers the foundation of that stack — the physical field layer through the data pipeline that feeds the twin.
1. Physical Field & Control Layer
PLC/SCADA systems, smart sensors (vibration, temperature, pressure), actuators, and embedded controllers (microcontrollers, FPGA).
2. Edge Processing & Gateway
Protocol normalization (OPC UA → MQTT), edge inference and aggregation, local store-and-forward caching, and device authentication (mTLS).
3. Data Pipeline & Storage Layer
High-throughput ingestion into time-series stores (InfluxDB/Timescale), spatial data (PostGIS), event streaming (Kafka/Pulsar), and object storage for CAD/BIM assets.
4. Twin Processing & Analytics Engine
Physics-based simulation and ML-driven anomaly detection, extended with LangGraph reasoning agents in my LLM-powered Digital Twin project.
5. Visualization & Operational UI
Operator-facing dashboards and natural-language interfaces that surface twin state, diagnostics, and recommendations in real time.
Featured Project: LLM-Powered Digital Twin
A real-time Digital Twin architecture that integrates IoT sensor data with rule-based system modeling and an LLM-powered reasoning agent.
- Live Digital Twin state modeling & Health evaluation
- Natural language interaction with physical systems
- Explainable diagnostics and recommendations
- RAG-powered reasoning & LangGraph-based orchestration
Pillar 2: Edge AI — Intelligence Where the Data Is Born
Cloud round-trips are too slow, too costly, and too fragile for physical systems that must react in milliseconds. I deploy quantized models directly onto MCU-class hardware, so detection and decisions happen on the device — online or off.
Industrial Edge AI Gateway
GRU-based inference on a dual-core microcontroller with TFLite Micro and FreeRTOS isolation — real-time anomaly detection for industrial safety with sub-100ms inference latency.
On-Device Inference
Quantized TFLite models for real-time sensor telemetry on MCU-class hardware — model size, latency, and power treated as design constraints, not afterthoughts.
Edge AI or Traditional Algorithms?
Rule-based logic still wins many edge problems. A practical comparison of when deterministic algorithms beat quantized models — and when they don’t — from real deployments.
The Cognitive Layer: LLM Systems
LangChain, LangGraph, and RAG are the reasoning layer over the two pillars — the part that explains what the physical system is doing and why. They support the edge and the twin; they do not replace physics or control logic.
LangChain & RAG
Context-aware reasoning over live system state, vector databases (FAISS, Chroma) for technical documentation, and grounded retrieval.
LangGraph
Stateful multi-node agent workflows, event-driven reasoning loops, and Digital Twin agents with persistent memory.
Digital Twin Agents
Agents that sit on top of the twin and answer the operator’s real questions — what is happening, why it is happening, and what should be done next.
Technical Stack
Edge & Embedded
- Hardware: Microcontrollers, embedded processors, FPGA
- OS: Embedded Linux, FreeRTOS, Bare Metal
- ML: TensorFlow, TFLite (Quantization), Scikit-learn
Physical Systems & IoT
- Modeling: Digital Twins, Edge AI Frameworks
- Protocols: MQTT, Modbus, BACnet, TCP/IP, SIP
Cognitive Layer — AI & LLM
- Frameworks: LangChain, LangGraph
- RAG: FAISS, Chroma, Vector Pipelines
Cloud & Platform
- Infrastructure: AWS (IoT Core, Lambda), Azure
- Backend: FastAPI, Django, Spring Boot
- Storage: Redis, InfluxDB, PostgreSQL, MongoDB
- Architecture: Microservices, Event-Driven
- Container: Docker, Kubernetes basics
Career Highlights
- [Edge AI] Physical AI Framework (Epic Safety): Architected an Edge AI framework on microcontrollers using quantized TFLite models for fire suppression, achieving sub-100ms inference latency for industrial safety.
- [IoT Platform] IoT Device Management (VTech): Engineered a global platform on AWS IoT Core, leading cross-functional teams to design high-performance device-side agents.
- [Scale] Mass Deployment (NTT): Delivered 200K+ Android tablets to NTT Japan; architected custom home launchers and SIP/RTP multimedia frameworks.
- [Embedded] Multimedia SDK Pioneers (Varovision): Developed MPEG-2 TS demuxers and motion detection algorithms for early mobile DTV and portable media players.
Apps I Built for My Own Life
The ideas are mine alone — I built these apps because I needed them in my own daily life. Everything after the idea was run like a real product project: detailed development plans and documentation in Confluence, project management through Jira tickets, and testing and debugging carried out with AI agents and coding assistance.
We have entered an era where the apps you need are apps you can build yourself. These two are my proof — born at my kitchen table, shipped to Google Play.
SmartShopper
Offline grocery lists with auto-categorization and easy management. No account, no connection — just a list that organizes itself by category while you shop.
Shopora
Shopping lists, loyalty & membership cards, and receipts with on-device OCR — plus monthly budgets and PDF/CSV export, all private on your device.
Architectural Philosophy
- LLMs as Cognitive Layers: Language models should reason, explain, and assist—not replace physics or control logic.
- State Over Data: Digital Twins must model behavior and evolution, not just metrics.
- Production First: Observability, tracing, and evaluation are mandatory for real-world AI systems.
- Human-Centric AI: Systems should communicate clearly with operators and engineers.