Toloka provides high-quality training data for AI models and agents by combining human expertise with AI-powered data annotation and crowdsourcing. Pulse 2.0 interviewed Toloka CEO Olga Megorskaya to learn more.
Olga Megorskaya’s Background
Screenshot
Could you tell me more about your background? Megorskaya said:
“I studied Mathematical Modelling in Economics and have been dealing with statistics, ML, and human-in-the-loop systems from the very beginning of my career. From the start, I’ve been focused on one core mission: building reliable ML and AI systems through smart human-AI cooperation. I’ve spent years scaling human-in-the-loop systems that let organizations scale ML and AI development without sacrificing quality.”
“Before founding Toloka, I led work to build data production infrastructure and scale crowdsourced labeling for machine learning products across industries like IT, e-commerce, telecom, and ride tech. That experience gave me a deep foundation in both the technical side—especially NLP and computer vision—and the operational reality of what it takes to deliver high-quality AI at scale.”
Formation Of The Company
How did the idea for the company come together? Megorskaya shared:
“Toloka was founded in 2014 initially as an internal project within a large corporation because we saw a clear scaling problem in ML development: high-quality data labeling simply couldn’t keep up with the pace and volume machine learning demanded. The early insight was that you need a flexible, distributed crowd to generate and verify data at industrial scale—not a purely in-house workflow. We built Toloka to turn human insight into scalable infrastructure: a platform where large, complex annotation and evaluation tasks could be broken down, routed to the right contributors, and quality-controlled systematically. Today, we’ve evolved from the simple tasks of the early days to generating high-complexity, highest-quality data fueling frontier models and AI agents, but our core philosophy remains: managing human efforts at scale in a technological way.”
Favorite Memory
What has been your favorite memory working for the company so far? Megorskaya reflected:
“My favorite memory is not a single event, but a specific recurring realization I keep having. For years, we focused on “ground truth” for classical ML. When LLMs arrived, there was a moment where the industry feared human data would become obsolete. Instead, the opposite happened – the industry realized very quickly how critical human data is to the success of AI. It validated our long-term thesis: that as AI gets smarter, the human input doesn’t disappear; it just becomes more nuanced and specialized. Seeing our infrastructure successfully evolve from simple labeling to complex expert reasoning, to now developing RL gyms, was a pivotal moment for the team.”
Core Products
What are the company’s core products and features? Megorskaya explained:
“Our ecosystem is built on two core pillars: our proprietary LLM QA technology—which orchestrates and verifies human effort—and our global marketplace of experts.”
“On this foundation, we have built three distinct products:
Managed Service Data Business: For our enterprise and lab partners, our experienced team of engineers and researchers builds complex, custom data solutions. We act as a strategic partner to handle the heaviest technical lifts, including designing RL Gyms, running complex model evaluations, and managing large-scale annotation projects.
Self-Service Data Annotation Platform: This platform is designed for researchers, enterprises and startups who need to experiment fast. It provides direct access to our crowd of experts and our framework of quality management, allowing users to iterate on data annotation independently and quickly.
Tendem: This is our hybrid AI agent for the prosumer market. It allows users to delegate complex tasks—like deep-dive research or data scraping at scale. Unlike standard AI tools, Tendem doesn’t just generate a draft; it routes the work to a human expert who verifies facts, fixes nuances, and ensures the output is client ready. It essentially packages our industrial-grade infrastructure into a simple chat interface, giving individuals the reliability of a managed team without the overhead.”
“Together, these three offerings cover the whole spectrum of tasks that involve humans for AI—from training the models to applying them in the real world.”
Challenges Faced
Have you faced any challenges in your sector of work recently? Megorskaya acknowledged:
“The last five years have been challenge after challenge. From lightning-fast technological advances that not only threatened our existence but forced us to question our core value proposition, to geopolitical events that impacted our teams, we’ve had to redefine, reimagine, and reinvent Toloka a few times over. We overcame this by staying close to the technical reality of the market. When LLMs commoditized simple labeling, we didn’t fight it; we moved up the value chain to “expert-in-the-loop” for complex reasoning. I’m incredibly proud of how far our team has gone, and today we are partnering with industry leaders and solving the industry’s most complex challenges.”
Evolution Of The Company’s Technology
How has the company’s technology evolved since launching? Megorskaya noted:
“When we started, our technology was built for simple, repetitive micro-tasks—identifying objects in images or classifying search queries—using a general global crowd. As model capabilities advanced, we expanded and evolved to handle high-complexity, multi-turn reasoning tasks by adding a curated network of domain experts like physicists, developers, and writers.”
“Most recently, we have moved into developing entire virtual environments (RL Gyms) for training and evaluating AI agents. However, it is important to understand that each new layer does not cancel the previous ones. For example, while we are building complex expert systems today, there’s a massive resurgence in demand for our large, diverse crowd of annotators to support the rise of Robotics, where large-scale, diverse data collection is critical.”
“The launch of Tendem represents the final step in this evolution: taking this industrial-scale infrastructure—orchestration, expert escalation, and layered verification—and wrapping it into a single, accessible agent experience for prosumer audiences.”
Significant Milestones
What have been some of the company’s most significant milestones? Megorskaya cited:
“The past year has been defined by three major pillars of progress: pioneering research, strategic funding, and product expansion. First, Toloka led pioneering work in the development and evaluation of AI agents, particularly with our RL gym environments. We’ve moved beyond static data to dynamic evaluation, setting a new standard for how safe and reliable agents are built. Second, we raised a $72M funding round.. This capital is fueling the build and roadmap for Tendem, while simultaneously accelerating the development of Toloka’s core data products. Third, we executed two major product launches that expanded access to our data infrastructure.”
- The Toloka Self-Service Platform: We launched our new self-service platform to give research teams, startups and enterprise teams access to our network to generate high-quality expert data. With our AI Assistant handling project setup and proprietary LLM QA validating every output, we’ve made it possible for anyone to get quality data for AI models and agents with transparent pricing.
- Tendem: Parallel to that, we launched Tendem, our hybrid agent for the prosumer market. We are actively onboarding users across marketing, sales and consulting, with thousands of users helping us validate that the hybrid model successfully bridges the reliability gap left by pure-AI agents.
Customer Success Stories
Can you share any specific customer success stories? Megorskaya highlighted:
“ While most of our work is under NDA, I can share the impact of our infrastructure. For example, we worked with Poolside, a company building AI for software development. They used our platform to access vetted developers who create complex coding problems and solutions to train their models. On the enterprise side, we worked with Shopify to build the master product catalog powering their AI features. Faced with mapping millions of noisy products to a 10,000-label taxonomy, we built a hybrid pipeline where AI handled the structure and humans resolved the edge cases—achieving 95% accuracy in days. We see similar efficiency with Tendem users who are offloading deep-dive market research tasks that previously took a week to complete and verify and receive a client-ready brief in a fraction of the time.”
Total Addressable Market (TAM)
What total addressable market (TAM) size is the company pursuing? Megorskaya assessed:
“We are operating at the intersection of two massive markets. First is the AI data services market, which is projected to reach over $10 billion by 2030 as demand for specialized RLHF and expert data grows. Second, with Tendem, we are addressing the broader knowledge-work automation market. By combining the speed of AI with the reliability of freelance talent, we are targeting a segment of the trillion-dollar global freelance and agency services economy—specifically the portion of work that requires human judgment but is repetitive enough to be orchestrated by AI.”
Differentiation From The Competition
What differentiates the company from its competition? Megorskaya affirmed:
“Our core differentiator is our technological approach to quality control. While competitors rely on operational management, we treat quality as an engineering challenge. We rely on the Toloka Quality Loop, a system where quality is defined, measured, and verified through automated infrastructure and AI agents.”
“Our core quality infrastructure underpins everything we do. Whether it’s complex reasoning data for LLMs, reliable outputs for Tendem users, or our new self-service platform, we use agentic workflows and a mix of LLM-driven and human QA to guarantee quality. By replacing manual oversight with engineering, we deliver higher accuracy and consistency at a scale that traditional operations cannot sustain.”
Future Company Goals
What are some of the company’s future goals? Megorskaya emphasized:
“We have already built an ecosystem where human-AI collaboration covers the entire lifecycle of technology—from training models to applying them. Our goal now is to significantly expand and deepen this access to bring more value to a wider range of users.
We are scaling our investment in all three layers of our ecosystem to grow our impact:
- For Enterprise: We are continuing to move upstream into the most complex post-training challenges, enabling more labs to deploy custom RL Gyms and evaluation pipelines that standard vendors cannot handle.
- For Developers: We are aggressively expanding our self-service capabilities to ensure that any researcher or startup can iterate on data as fast as they iterate on code, without the friction of traditional vendor management.
- For Individuals: With Tendem, we are growing our reach in the prosumer market, helping more individuals solve one-off, complex tasks where AI alone currently hits a wall.
This expansion drives our “expert economy.” By growing all three channels, we bring a wider variety of work to our global expert network, providing them with more sustainable opportunities to monetize their skills while giving our customers—big and small—continued access to the best minds.”
Additional Thoughts
Any other topics you would like to discuss? Megorskaya concluded:
“I think it is important to touch on the future of work. We believe the future isn’t about AI replacing humans, but about infrastructure that allows humans to move up the stack. With Tendem and Toloka, we are building a system where AI handles the ‘blank page’ problem and the rote execution, allowing human experts to act as editors, strategists, and verifiers. This doesn’t just make businesses more efficient; it creates a more sustainable model for digital work where contributors are paid for their expertise and judgment rather than repetitive labor.”
