Let’s be real, most AI projects die at the prototype stage. Great demo, cool tech, zero production-readiness. What separates those graveyard prototypes from systems that actually run the business? We’d argue it’s people like Andrei Shcherbinin, Team Lead at Social Discovery Group, who’s quietly been building ML infrastructure that does serious heavy lifting. t Social Discovery Group Social Discovery Group We grabbed some time with him to talk about engineering roots, 12x speed improvements, chatbots that handle 95% of support on their own, and why he thinks knowing your math is still the best career move in tech. 🗣️ You didn’t take the bootcamp-to-big-tech route. PhD ABD, deep signal processing background — does any of that actually matter when you’re building ML systems today? You didn’t take the bootcamp-to-big-tech route. PhD ABD, deep signal processing background — does any of that actually matter when you’re building ML systems today? You didn’t take the bootcamp-to-big-tech route. PhD ABD, deep signal processing background — does any of that actually matter when you’re building ML systems today? — More than people think. Engineering school doesn’t just teach you formulas — it teaches you to think in systems. When I was working with signal processing, the whole game was extracting clean, useful information from enormous streams of noisy data in real time. Turns out? That’s basically what a recommender system or attribution algorithm does. Different tools, same problem. The other thing a serious academic background gives you is discipline around experimentation. In science, one successful run doesn’t mean anything. You need rigorous methodology. In high-load systems, that same mindset is what separates something that scales from something that collapses the moment traffic spikes. 🗣️ Let’s talk about the attribution model overhaul. Marketing attribution sounds… dry. Why was it actually a hard engineering problem? Let’s talk about the attribution model overhaul. Marketing attribution sounds… dry. Why was it actually a hard engineering problem? Let’s talk about the attribution model overhaul. Marketing attribution sounds… dry. Why was it actually a hard engineering problem? — Ha, it only sounds dry until you realize “which ad channel actually drove this purchase” is a question worth millions of dollars. Standard attribution models take a lazy shortcut — they give all the credit to the last click. Reality is messier. Users interact with a brand across multiple touchpoints before buying, and you need probabilistic models to figure out how to fairly distribute credit across all of them. The real pain wasn’t the math, though. It was the time. Our existing calculations took six hours to run. Marketing can’t wait six hours — they need to know what’s working now. So we rebuilt the algorithms and the data processing architecture from scratch. time six hours to run now Result: calculation time dropped from 6 hours to 30 minutes. That’s a 12x improvement. And because we now had accurate attribution, we stopped spending budget on channels that looked good under the old model but were actually doing nothing. That math converts directly into saved money. Result: Result: calculation time dropped from 6 hours to 30 minutes. That’s a 12x improvement. And because we now had accurate attribution, we stopped spending budget on channels that looked good under the old model but were actually doing nothing. That math converts directly into saved money. 30 minutes looked 🗣️ 12x faster means the infrastructure underneath had to change too, right? 12x faster means the infrastructure underneath had to change too, right? 12x faster means the infrastructure underneath had to change too, right? — Completely. We moved from manual process management to fully automated pipelines — MSSQL and BigQuery feeding into Airflow and Kafka, through to Athena and S3. Before that, a lot of processes needed human intervention, which introduced both delays and human error. After the automation: manual workload down 80%, data latency reduced by 35%. And crucially, the architecture scales — when load increases, the system just uses more resources instead of falling over. manual workload down 80% 35% 🗣️ The support chatbot is another wild one. 95% of conversations automated, accuracy of 0.95+. How? The support chatbot is another wild one. 95% of conversations automated, accuracy of 0.95+. How? The support chatbot is another wild one. 95% of conversations automated, accuracy of 0.95+. How? — The core problem with most chatbots is context blindness. We built a model that recognizes 35 distinct user intents — so it knows the difference between someone asking for a refund and someone with a technical issue, and responds appropriately. 35 distinct user intents But the part I’m genuinely proud of isn’t just the accuracy — it’s the security layer. We work with large language models, and you simply cannot feed raw user data into them. So before any message hits the neural network, it goes through an anonymization filter. Personal data stripped, then processed. Users get fast, intelligent responses; their privacy stays intact; we stay compliant. It’s a security gateway built into the core architecture, not bolted on afterward. 🗣️ You’ve made model monitoring a standard across your ML ecosystem. Why is that a bigger deal than it sounds? You’ve made model monitoring a standard across your ML ecosystem. Why is that a bigger deal than it sounds? You’ve made model monitoring a standard across your ML ecosystem. Why is that a bigger deal than it sounds? — Because of something called data drift. A model trained on last month’s user behavior can quietly start making bad decisions this month — and without monitoring, you might not notice for weeks. You’re just losing money while the algorithm degrades and nobody’s looking. weeks Our monitoring system surfaces problems in real time. If recommendation quality dips or classification accuracy drops, engineers get notified immediately, and not via angry user tickets. We cut problem detection time by 60% and reduced model retraining prep time by 40% because the system automatically flags what’s changed and needs attention. 60% 40% It turns ML maintenance from constant firefighting into something you can actually plan and manage. 🗣️ You manage a cross-functional team now — ML engineers, MLOps, backend devs. How do you get them to actually work well together? You manage a cross-functional team now — ML engineers, MLOps, backend devs. How do you get them to actually work well together? You manage a cross-functional team now — ML engineers, MLOps, backend devs. How do you get them to actually work well together? — The classic failure mode is communication gaps. An ML engineer builds something great; a backend dev has no idea how to integrate it. You end up with handoff chaos. We fixed that with a clear responsibility matrix and KPIs — everyone knows their scope and who to go to for what. We also standardized documentation and onboarding. New hires used to spend weeks just trying to understand the architecture. Now that it’s all documented properly, onboarding time is down 40%. onboarding time is down 40% Managing a team is genuinely similar to managing a distributed system: reduce uncertainty, define interfaces clearly, and things get faster and more reliable. 🗣️ Last one: what should engineers focus on if they want to build things at this level? Last one: what should engineers focus on if they want to build things at this level? Last one: what should engineers focus on if they want to build things at this level? — The trend is toward easier model access but more complex orchestration. MLOps is becoming its own discipline, and the engineers who understand the full stack — cloud, databases, security, and the models — are going to be the most valuable people in the room. and But honestly, the thing that hasn’t changed and won’t change is the fundamentals: statistics, linear algebra, algorithmic thinking — these are constant. Tools change every six months. The principles don’t. Learn those, and you can adapt to anything. Tools change every six months. The principles don’t. Also: always ask what the business impact is. The attribution project wasn’t about doing fancy math — it was about optimizing a marketing budget. When your team understands that, the quality of every decision they make goes up. Also: always ask what the business impact is. Also: always ask what the business impact is. The attribution project wasn’t about doing fancy math — it was about optimizing a marketing budget. When your team understands that, the quality of every decision they make goes up. Andrei leads the ML team at Social Discovery Group. If you want to geek out about ML infrastructure, data pipelines, or what it actually takes to get a model to production — join our team! Find SDG Careers — https://socialdiscoverygroup.com/vacancies/ Andrei leads the ML team at Social Discovery Group. If you want to geek out about ML infrastructure, data pipelines, or what it actually takes to get a model to production — join our team! Find SDG Careers — Social Discovery Group. Social Discovery Group. https://socialdiscoverygroup.com/vacancies/ https://socialdiscoverygroup.com/vacancies/

This story contains new, firsthand information uncovered by the writer.

Discovery

From Satellite Signals to Neural Networks

About Author

Comments

TOPICS

THIS ARTICLE WAS FEATURED IN

Related Stories

AI in Customer Success: How to Balance Automation and Personalisation

Bright Data vs Oxylabs: Which Is Better According to Customers?

From MQTT to SQL: A Practical Guide to Sensor Data Ingestion

The Internet Is Drowning in “Thought Leadership”

The Noonification: How Often Do NFTs Pass The Howey Test? (1/13/2023)

Darwin's Hybrid Intelligence to Align AI & Human Goals for Startups & VCs

AI in Customer Success: How to Balance Automation and Personalisation

Bright Data vs Oxylabs: Which Is Better According to Customers?

From MQTT to SQL: A Practical Guide to Sensor Data Ingestion

The Internet Is Drowning in “Thought Leadership”

The Noonification: How Often Do NFTs Pass The Howey Test? (1/13/2023)

Darwin's Hybrid Intelligence to Align AI & Human Goals for Startups & VCs

Light-Mode

Classic

Newspaper

Minty

Dark-Mode

Neon Noir

Minty

HN StartUps