#040 ML Александр Дончук. Про стартапы, бизнес и ML в промышленности

#040 ML Александр Дончук. Про стартапы, бизнес и ML в промышленности

#040 ML Alexander Donchuk. About startups, business and ML in industry

by Machine Learning Podcast · View the channel on PSI

Applied Sciences & Engineering AI — Models & Research UYQArtificial intelligence (AI)UYQMMachine learning
🎙 Machine Learning Podcast 👥 5K 📅 September 20, 2022 ⏱ 62 min 👁 476 📄 expert opinion 🧭 2026-10-09
Available in: English (current) Français

Keywords

defectoscopyrailwaycomputer visionindustrial AIstartup journey

Summary

In this episode of the Machine Learning Podcast, host Mikhail interviews Alexander Donchuk, co-founder of Aurora AI, about applying machine learning and computer vision to industrial defect detection, specifically in the railway sector. Alexander shares his journey from working on electric vehicle charging startups to discovering ML in 2017 and founding a company that develops automated inspection systems for freight train components. The discussion covers the challenges of scaling industrial AI solutions, the initial reluctance of industry players, and the decision to focus on brake pad inspection and crack detection. He describes their first prototype using high-speed cameras and custom lighting, achieving more precise measurements than human inspectors. The conversation delves into dataset collection, annotation, and a multi-stage pipeline involving segmentation masks, anomaly detection, and crack detection. Alexander also discusses the importance of accelerators like Sber 500 and the role of luck in entrepreneurship. The episode provides practical insights into the complexities of deploying ML in conservative industries and the business realities of industrial startups.

165 words

Critical Evaluation

Value of the Information & Strength of the Argument

The value of the information lies in the detailed, firsthand account of implementing computer vision in a highly regulated and conservative industry. Alexander provides concrete technical details, such as camera setup, speed constraints, and measurement accuracy, which are useful for practitioners considering similar applications. The argumentation is coherent and evidence-based, drawing from his direct experience with deploying systems in real-world conditions, including extreme temperatures. The discussion on why industrial ML projects are difficult to scale (due to niche per-case solutions) is insightful and well-articulated.

Scientific Rigor, Source Quality, Title Accuracy

The video is an interview, not a scientific study, and thus lacks formal citations. However, the guest references his own company’s work and experiences, which are self-reported and not independently verified. The title matches the content accurately, focusing on ML applications in industry and startup aspects. The description includes links to the podcast episode, related resources, and a book on customer development, but these are not directly cited during the conversation. The absence of external third-party sources limits the scientific rigor, though the practical nature of the content compensates somewhat.

189 words

Title / Content Match

The title accurately reflects the content: the episode discusses applying ML in industry, specifically railway defectoscopy, along with startup and business aspects.

Quality & Reliability

6/10

The podcast is an interview with a practitioner sharing experiential insights about applying machine learning to industrial defect detection. While it offers concrete technical details and real-world examples, it lacks formal citations or empirical validation, and the reliability is based on anecdotal evidence rather than peer-reviewed sources.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

Contribution & Novelties

This episode offers original insights into the practical challenges of applying computer vision in industrial settings, particularly the railway sector, where conservative practices and harsh environments complicate deployment. The guest’s detailed account of building a defect detection system from scratch, including dataset collection and pipeline design, provides valuable lessons for practitioners. The discussion on why industrial AI projects often fail to scale due to niche use cases adds a nuanced perspective on the business of applied ML.

Pour aller plus loin :

  • Computer vision — Overview of the field central to the episode’s technical approach.
  • Predictive maintenance — Related concept for using data to anticipate equipment failures, relevant to the inspection use case.
  • Nondestructive testing — General methodology for inspecting materials without damage, aligning with the brake pad inspection discussed.

130 words

Radar Profile

The radar profile shows moderate scores across all dimensions, with quantity of information slightly higher than quality and reliability. This suggests the content provides a substantial amount of practical detail but lacks rigorous sourcing and formal verification, typical of an expert-interview format rather than a peer-reviewed presentation.

Reliability 5/10