About

Building Reliable AI for Enterprise and Harsh Industrial Environments

Hi, I am Dmitrii Iunovidov, PhD.

I bridge the gap between theoretical machine learning and physical industrial production. After spending over 10 years in data-driven development and chemical industry R&D, I saw a recurring problem: theoretical AI often fails on the factory floor or in real-life applications. Heavy neural networks require expensive GPUs, cloud dependencies create security risks, and LLM hallucinations are unacceptable in high-stakes environments.

My mission is to build autonomous, reliable, and closed-loop AI systems that actually work in harsh, real-world conditions.

What I am building right now

Currently, I am the Founding ML Engineer at LogicYield, a startup developing autonomous edge AI for the chemical and fertilizer industries. We digitize complex inert processes to eliminate defective products, optimize operations, and reduce profit losses.

Our core industrial systems include:

• DotPulse: An automatic optical particle size analysis system. It replaces slow laboratory checks (which typically happen only once every 4 hours) with real-time granulation control directly on the conveyor belt. This allows for instant response to deviations and saves significant costs on raw materials and energy.
• GuardDetector: A digital control system for industrial rounds and occupational safety. It uses advanced video analytics to track actual equipment inspections and PPE compliance. The system runs entirely on edge devices, reducing the Total Cost of Ownership (TCO) and ensuring high information security.
• Cognitive Hub: An AI system that organizes complex operational data into an intuitive communication interface for plant management and process operators, powered entirely by local CPU-based LLMs and NER engine.

My Technical Approach

To make this possible, we cannot rely on standard heavy AI architectures. My research and engineering work focuses on optimizing AI for strict CPU limits while maintaining deep semantic understanding.

• Edge Computer Vision: I am actively working on reformulating spatial instance segmentation and temporal tracking as a 1D sequence-tagging problem. By mapping visual streams into an IOB-2 grammatical framework using MobileNetV3 and BERT encoders, we can process complex environments without expensive 3D components or heavy post-processing.
• Enterprise LLM Orchestration: I design RAG and LLM systems that prioritize strict data retrieval. By using local structured databases and edge processing, we eliminate hallucinations and ensure privacy by design.

Let's Connect

Whether you are an industrial plant manager looking to deploy a pilot zone for real-time production control, a venture capitalist interested in deep-tech industrial AI, or a technical team looking for expertise in AI architecture, I am open to discussion.

My stack includes

Coding:

MLOps:

Hardware:

The copyright for site materials are distributed under Creative Commons Attribution-ShareAlike CC BY-SA license (if not said otherwise).