# Agajan Torayev Founder, engineer, and product builder. I am an AI founder, machine-learning engineer, robotics researcher, and software engineer based in London. I work on agentic AI systems for the physical world: video intelligence, industrial robotics, manufacturing optimization, AI-native CAD, product-design workflows, and practical tools that connect human intent with machine execution. I am the Co-founder and CEO of [RoboMinder](https://www.robominder.ai/), a UK AI company building agents for physical and digital workflows. Today, I lead the company across product, technical direction, commercial strategy, and fundraising. My background combines startup execution with PhD-level research in manufacturing engineering. Before founding RoboMinder, I worked as a Marie Sklodowska-Curie Early-Stage Researcher at the University of Nottingham, where my PhD research focused on intelligent decision-making and manufacturing-system configuration for Industry 4.0. ## Favourite Quote > The bitter lesson in 26 words: > Don't be distracted by human knowledge, as AI has been historically. > Instead focus on methods for creating knowledge that scale with computation, like search and learning. > — Richard Sutton ## What I Work On The throughline in my work is translation. I like building systems that take messy real-world input, incomplete human goals, or hard-to-use technical data and turn them into something structured, inspectable, and useful. That has meant: - Turning sketches, product ideas, and requirements into CAD-ready design artifacts - Turning video streams into searchable operational knowledge - Turning robot and process data into monitoring, diagnostics, and optimization decisions - Turning manufacturing requirements into system-configuration decisions - Turning static-analysis findings into verified security reports - Turning research methods into usable software products I am especially interested in AI systems that do useful work under real constraints: physical machines, production lines, product geometry, safety, downtime, tolerances, manufacturing cost, and the need for expert review. ## Current Work ### RoboMinder I co-founded and lead [RoboMinder](https://www.robominder.ai/), where we build AI agents for physical and digital workflows. We began with industrial computer-vision monitoring and have evolved into a broader agentic-AI company. Our work connects video understanding, robot and machine analytics, operational event detection, physical-operations intelligence, and AI-native product-design workflows. As CEO and co-founder, I work across company strategy, product direction, technical roadmap, customer discovery, pilots, commercial partnerships, fundraising, investor relations, sales, hiring, and team building. I enjoy working across those boundaries because it keeps the technology grounded in real customer workflows. Our enterprise customer context includes Boots, Chippindale, Morrisons, and Butternut Box. Reeco is a public automation partner and customer context for our palletising analytics work. Our investor and backer context includes Antler, Fuel Ventures, and Hatcher+. Our public grant/support context includes Innovate UK. A concrete example is the RoboMinder-powered AI palletising software and analytics with Reeco: live robot reporting, production-line insight, event logging, downtime and idle-time visibility, utilization metrics, live video feed, and error-event playback for palletising operations. ### Video AI Our Video AI work is about turning raw video into operational knowledge. I describe it as "Video to Knowledge" and "the CTRL+F for your physical operations." The idea is simple: video should not only be a passive recording. It should become searchable, queryable, event-aware, and useful for operators, engineers, managers, and operations teams. The environments I think about include: - Manufacturing production lines - End-of-line automation - Palletising and cobot systems - Robot cells - Warehouse and logistics operations - Retail operations - Quality-control workflows - Physical asset monitoring In practice, we combine video ingestion, event detection, computer vision, multimodal AI, analytics, dashboards, and human review. The real value is not only detecting something in a frame; it is helping people understand what happened, when it happened, why it matters, and what to do next. ### TexoCAD With [TexoCAD](https://www.texocad.ai/), we are building an AI product team for physical product builders. It helps turn sketches, ideas, and requirements into product specs, 3D models, CAD files, and simulation-ready designs. I do not think of TexoCAD as a simple "text prompt to 3D model" product. I think of it as an agentic product-development workflow. I structured TexoCAD around specialized roles: - Requirements engineering: briefs, scope, product requirements, and acceptance criteria - Concept engineering: ideas, sketches, visual direction, and early 3D concepts - CAD engineering: parts, CAD, DFM, STEP output, and editable geometry - Simulation engineering: FEA, process thinking, robotics context, and risk assessment - Human expert review: manufacturability, tolerances, materials, safety, fit, durability, and real-world engineering judgment The motivation behind TexoCAD is personal and practical: many people can describe what they want to build, but CAD and product-design tools often create a gap between thought and object. AI agents can help draft geometry, generate variations, wire parameters, produce first-pass CAD, and automate repetitive modeling work. Human experts still matter for the high-judgment parts: manufacturability, tolerances, material selection, compliance, safety, durability, serviceability, and whether a design should actually be made. ## Career Timeline ### RoboMinder Ltd - Role: CEO and Co-founder - Location: London, United Kingdom - Dates: December 2023 to present I founded and lead RoboMinder from company formation to market-facing products. My work covers product strategy, AI and computer-vision architecture, customer discovery, commercial partnerships, fundraising, investor relations, sales operations, and the technical roadmap. Metrics and outcomes include: - Secured £500,000 in pre-seed funding at a £3.5M valuation from Antler UK, Fuel Ventures, and Hatcher+ - Built and led a cross-functional team of engineers and researchers focused on scalable AI-driven visual intelligence systems - Established strategic partnerships and commercial agreements with major retailers including Boots and Morrisons - Defined the company vision, product strategy, and go-to-market plan, reaching early product-market fit in retail technology - Led the technical roadmap, fundraising, investor relations, and sales operations across UK and international markets ### University of Nottingham, Institute for Advanced Manufacturing - Role: Marie Sklodowska-Curie Early-Stage Researcher and PhD Researcher - Location: Nottingham, United Kingdom - Dates: September 2020 to September 2023 At Nottingham, I worked on AI-based manufacturing-system configuration and intelligent decision-making for Industry 4.0. My PhD thesis was titled "A holistic methodology for manufacturing systems configuration." My work included Bayesian optimization for industrial robot energy consumption, reinforcement-learning-based decision-making for manufacturing configuration, modular software architectures for plug-and-produce manufacturing apps, and autonomy models for manufacturing systems. Outcomes from this period include: - Reduced industrial robot energy consumption by 33% - Improved manufacturing-system response time by 25% - Increased lab industrial robot utilization by 50% by developing Python tooling for industrial robot control - Built `fanucpy`, which has been used in more than 10 international research and commercial organizations ### University of Bonn, Computer Graphics Group - Role: Scientific Worker - Location: Bonn, Germany - Dates: November 2019 to April 2020 After my MSc, I worked on a DFG-funded medical visualization project related to propagator-based diffusion imaging data. My work focused on deep-learning-assisted interactive classification of diffusion MRI data using feature representations from a dual-branch CNN autoencoder. Metrics and outcomes include: - Increased the F1 score of diffusion MRI classifications by 50% - Reduced expert annotations by 2.08x - Published "Interactive Classification of Multi-Shell Diffusion MRI With Features From a Dual-Branch CNN Autoencoder" - Received Best Paper Honorable Mention recognition at the Eurographics Workshop on Visual Computing for Biology and Medicine ### Fraunhofer IAIS - Role: Research Intern - Location: Sankt Augustin, Germany - Dates: October 2018 to April 2019 At Fraunhofer IAIS, I worked on applied NLP and text-mining software. My contribution focused on improving inference performance of an internal text-classification tool through code restructuring and removal of unnecessary condition checks. Metric: - Decreased text-classification inference time by 15% ### Magtymguly Turkmen State University - Role: Teaching Professional - Location: Ashgabat, Turkmenistan - Dates: August 2015 to August 2017 I taught web programming, Python algorithms, and database management with MariaDB to first- and second-year university students. I also organized a computer-science club. Metrics and outcomes include: - Increased computer-science course attendance by 20% - Increased average student grades by 10% - Helped one computer-science club participant win first prize in a national project contest ### Central Bank of Turkmenistan - Role: Software Engineer - Location: Ashgabat, Turkmenistan - Dates: December 2012 to December 2013 I built a LAMPP-stack web application for automated analysis of daily reports from currency-exchange points. Metric: - Reduced daily office personnel work time by 1.5 hours This was an early example of the kind of software I still like building: automation that removes repetitive operational work. ### eSegment GmbH & Co. - Role: Software Engineering Intern - Location: Berlin, Germany - Dates: March 2012 to August 2012 I implemented a multi-threaded, robots.txt-friendly Python web crawler for data acquisition from e-commerce platforms. Metric: - Improved data acquisition speed by 75% ## Education ### PhD in Manufacturing Engineering - Institution: University of Nottingham - Institute: Institute for Advanced Manufacturing - Dates: 2020 to 2023 - Thesis: "A holistic methodology for manufacturing systems configuration" - Focus: manufacturing-system configuration, intelligent decision-making for Industry 4.0, reinforcement learning, Bayesian optimization, and autonomous manufacturing systems Research visits: - Mondragon University, Spain: reinforcement-learning-based optimal configuration selection for machining process planning - STIIMA-CNR, Milan, Italy: energy-consumption optimization methods for industrial robots using Bayesian optimization ### MSc in Computer Science - Institution: University of Bonn - Dates: 2017 to 2019 - Focus: intelligent systems, machine learning, deep learning, NLP, data mining, big data, and data science - Thesis: "Unsupervised Representation Learning for Volume Rendering Diffusion MRI" My MSc thesis received the best possible grade. ### Student Exchange in Electrical Engineering and Computer Science - Institution: Technical University of Berlin - Dates: 2011 to 2012 During this exchange, I worked at DAI-Labor, the Distributed Artificial Intelligence Laboratory, where I developed an open-source vulnerability scanner. ### Diploma in Applied Mathematics and Informatics - Institution: Magtymguly Turkmen State University - Dates: 2009 to 2015 - Specialization: applied mathematics for computer science - Diploma thesis: comprehensive analysis of the Viola-Jones object-detection framework and Haar features This degree gave me a foundation in mathematics, algorithms, programming, databases, probability, statistics, differential equations, complex numbers, and mathematical analysis. ## Research Themes ### Manufacturing Configuration and Decision-Making My PhD research focused on how manufacturing systems can select, configure, and adapt capabilities under changing requirements. This includes machines, tools, operations, process plans, constraints, and multi-criteria tradeoffs. ### Industrial Robot Optimization and Integration I worked on data-efficient optimization methods for industrial robots, especially energy consumption. This research required real robot experiments, not only simulation, which shaped how I think about AI systems that must work under physical constraints. ### Modular and Autonomous Manufacturing Software I contributed to work on plug-and-produce manufacturing apps, microservice-style architectures for manufacturing systems, and maturity models for autonomy. This connects factory systems with software engineering ideas such as modularity, APIs, interoperability, and containerization. ### Medical AI and Visualization At the University of Bonn, I worked on diffusion MRI classification using learned feature representations. This combined deep learning, random forests, medical visualization, and interactive workflows where human experts remain part of the loop. ## Selected Publications ### Online and Modular Energy Consumption Optimization of Industrial Robots - Venue: IEEE Transactions on Industrial Informatics - Year: 2023 In this work, I studied data-efficient optimization of industrial robot energy consumption using Bayesian optimization and real robot experiments. ### Towards Modular and Plug-and-Produce Manufacturing Apps - Venue: 55th CIRP Conference on Manufacturing Systems / Procedia CIRP - Year: 2022 In this paper, we explored manufacturing app architectures based on modularity, microservices, containerization, and communication technologies. ### A Maturity Model for the Autonomy of Manufacturing Systems - Venue: The International Journal of Advanced Manufacturing Technology - Year: 2023 In this work, we contributed to defining and evaluating autonomy levels in manufacturing systems. ### Optimal Selection of Manufacturing Configurations Using Object-Oriented and Mathematical Data Models - Year: 2023 In this work, we addressed manufacturing configuration selection using known equipment capabilities and capacities represented through object-oriented and mathematical data models. ### Interactive Classification of Multi-Shell Diffusion MRI With Features From a Dual-Branch CNN Autoencoder - Venue: Eurographics Workshop on Visual Computing for Biology and Medicine - Year: 2020 In this paper, we presented an interactive diffusion MRI classification workflow using learned feature representations. It received Best Paper Honorable Mention recognition. ## Open Source and AI Projects ### fanucpy [fanucpy](https://github.com/torayeff/fanucpy) is my most visible open-source robotics project. It is a Python interface for FANUC industrial robots and includes both Python-side robot interface code and FANUC controller-side driver code tested with FANUC R-30iB Mate Plus controllers. I built it to support robot connection, joint and cartesian motion, gripper control, robot state queries, external program calls, digital output, robot I/O access, and Python-driven robot application control. Why it matters to me: - It shows practical industrial robot integration experience beyond simulation - It provides a bridge between research software and factory hardware - It has public adoption from robotics and research users - It reflects my interest in tools that let engineers and researchers build faster ### Other Open-Source Projects - `mfgrl`: a manufacturing reinforcement-learning environment for configuration-decision research - `DeepMRI`: code related to learned feature representations for MRI analysis - `RoboCupHumanoid`: computer-vision work on soccer ball detection for the RoboCup Humanoid League - `digital-twins`: work related to digital twins for robotic assembly - `pyfactoryio`: a Python wrapper for Factory I/O ### Llama CCTV Operator I led the Llama CCTV Operator project at Meta's first LlamaCon Hackathon in San Francisco. The project used Llama 4 multimodal image understanding to help CCTV control-room operators define custom video events in natural language and detect those events in video chunks without fine-tuning. The architecture connected RTSP video streams, video chunking, multimodal event detection, backend services, event storage, and an operator-facing frontend. ### Codeminder Codeminder is a hybrid LLM-based SAST tool built for the Gemini Vibe Code Hackathon London. It combined Semgrep, Trivy, tree-sitter, dependency graphs, code slicing, Gemini-based vulnerability verification, FastAPI services, and actionable report generation. The idea was to combine deterministic static-analysis tools with LLM-based semantic verification instead of treating the model as an isolated text generator. ## Recognition - Third place at Meta's first LlamaCon Hackathon in San Francisco - Third place at Gemini Vibe Code Hackathon London with Codeminder - First place in a TECNALIA datathon for an oil-refining-process optimization project - Best Paper Honorable Mention at the Eurographics Workshop on Visual Computing for Biology and Medicine - Marie Sklodowska-Curie Actions ITN scholarship - DAAD Study Scholarship for Graduates of All Disciplines - ERASMUS / MANECA student exchange scholarship - PRINCE2 Foundation Certificate in Project Management ## Technical Focus ### AI Systems - Artificial intelligence, machine learning, data science, and applied AI product development - LLM agents, tool calling, structured outputs, and multi-step reasoning workflows - Multimodal AI for text, image, video, and structured data - Computer vision and video intelligence - RAG pipelines, embeddings, reranking, citation grounding, and memory systems - Human-in-the-loop review, guardrails, evaluation, and cost controls - OpenAI, Anthropic Claude, Google Gemini, Meta Llama, Vercel AI SDK, LangChain, LangGraph, LlamaIndex, PydanticAI, and Hugging Face workflows ### Machine Learning and Data Science - PyTorch, scikit-learn, pandas, NumPy, OpenCV, Hugging Face, RLlib, LightGBM, and XGBoost - Supervised learning, unsupervised learning, representation learning, classification, feature extraction, and model evaluation - Reinforcement learning, Bayesian optimization, model-based optimization, operations research, and multi-criteria decision-making - Data mining, NLP, big data, data science workflows, experiment design, and applied statistical thinking - Optimization tooling including Google OR-Tools, GEKKO, and pymoo ### Product Engineering - TypeScript, JavaScript, React, Next.js, React Server Components, Server Actions, Tailwind CSS, and shadcn/ui-style component systems - Python, FastAPI, Pydantic, REST APIs, streaming HTTP, WebSockets, and event-driven interfaces - Authentication and authorization patterns including JWT, OAuth, Clerk, and Auth.js - Git, GitHub, code review, CI, and automated quality checks ### Data and Production Ops - PostgreSQL, pgvector, Pinecone, Qdrant, Weaviate, ChromaDB, Redis, and SQL schema design - Vector search, hybrid search, semantic search, metadata filtering, and retrieval evaluation - OpenTelemetry, LLM tracing, prompt evaluation, dataset-based evaluation, golden tests, and regression suites - Docker, Linux, Vercel, Modal, Railway, Render, Fly.io, AWS/GCP concepts, queues, retries, scheduled workflows, secrets, monitoring, and release workflows ### Physical AI - Industrial robot programming and integration across FANUC, ABB, and KUKA contexts - ROS2, robot software architecture, robot application control, and robot-system integration - Simulation environments and digital manufacturing tools including MuJoCo, Isaac Sim, Visual Components, Factory I/O, and robot/digital-twin simulation workflows - Robot energy optimization, monitoring, analytics, palletising, and end-of-line automation - Manufacturing-system configuration, plug-and-produce manufacturing, digital twins, Industry 4.0, and autonomous manufacturing systems - CAD/3D workflows, STEP/CAD files, text-to-CAD, AI-native product design, simulation-ready design, and video-to-knowledge systems - Security and code-analysis workflows using Semgrep, Trivy, tree-sitter, dependency graphs, code slicing, and LLM-assisted vulnerability verification ## How I Think About AI I am most interested in AI systems that work in constrained environments where outputs have to be checked, modified, simulated, manufactured, deployed, or trusted. The systems I like building combine models with tools, structured outputs, domain constraints, human review, and feedback from the real world. I am less interested in isolated text generation and more interested in systems that help people move from intent to execution. In physical domains such as manufacturing, robotics, CAD, and operations, I am closer to Richard Sutton's view: the long-term leverage comes from scalable methods for creating knowledge, especially search and learning. Domain expertise is still useful for setting objectives, constraints, interfaces, and evaluation, but I do not want systems that simply encode human assumptions. I want systems that can explore, learn, test, and improve beyond the limits of hand-designed knowledge. ## Public Links - Website: [torayeff.com](https://torayeff.com/) - RoboMinder: [robominder.ai](https://www.robominder.ai/) - TexoCAD: [texocad.ai](https://www.texocad.ai/) - GitHub: [github.com/torayeff](https://github.com/torayeff) - X: [x.com/torayeff](https://x.com/torayeff) - Google Scholar: [my Google Scholar profile](https://scholar.google.com/citations?user=tiuDR1IAAAAJ) - LinkedIn: [linkedin.com/in/torayeff](https://linkedin.com/in/torayeff) - ORCID: [0000-0002-8141-5704](https://orcid.org/0000-0002-8141-5704) ## Freshness Note Some public metrics, such as GitHub stars, forks, citation counts, repository counts, and profile statistics, change over time. I prefer to refresh those numbers before using them in a pitch deck, investor update, grant application, formal CV, or claim about current status.