Toloka

We started Tendem by Toloka as a bold experiment: can we build a system to manage human efforts in solving ANY task? Specifically, unpredictable, one-of-a-kind, never seen before tasks? In our data business, projects are large and tasks are repeatable. You improve quality through training, expert selection, and iteration. But a task that's never appeared before and will never repeat again? That's a challenge we wanted to solve. A powerful AI agent to decompose and orchestrate the task. An expert network to handle the parts that need real judgment and human intelligence. Quality checks before anything reaches user. That's Tendem by Toloka . Since our first beta, the team has made enormous progress: more domains, more task types, expanded agent capabilities, and MCP so agent developers can call human expertise directly from their workflows. A universal human layer, callable by users or agents, wherever expert judgment is needed. That's where we're headed. Super proud for the team to be launching on Product Hunt today. Please check it out and support us: https://lnkd.in/eFVwWCQg

The architecture has changed significantly since our first release. Earlier builds handled the full task end-to-end. What’s live now is a major update that moves Tendem to a true hybrid pipeline: AI handles the structured, repeatable parts at speed, while vetted experts take on the parts that require human judgment, context, or domain experience — and then verify, correct, and sign off on the final result. The result: tighter handoffs, better outputs, and less wasted effort on both sides. If you tried Tendem by Toloka before, it’s worth another look. If you haven’t, today is a good time to start! Use PRODUCTHUNT to get $50 off your first task. We’re live on Product Hunt today and would genuinely appreciate your support. Link in the first comment ;)

Meet Toloka in Rio at #ICLR with something interesting on Benchmarks

Multimodality is on the rise again. This week's #Tolokacoolproject is about speech recording by non-native English speakers. Many of us have been there: you ask Siri or Alexa something and it completely misses the mark. Or you're on the phone with an automated agent and it doesn't quite understand your response. Getting voice recognition right across languages, dialects, and accents is genuinely hard. One team used Toloka 's self-service platform to collect a pronunciation dataset covering A–Z alphabet and 13 command words, sourced from verified non-native speakers across multiple countries and continents, using our Audio LLM QA to ensure the quality of recordings. The result: a project able to collect a clean, geographically diverse dataset, built without a custom recruitment pipeline, drawn from a contributor network spanning 100+ countries. Interesting that in such self-explanatory tasks additional quality measures such as task onboardings substantially reduce interested audience while not bringing equivalent quality uplift, which made us rethink our approach towards task onboardings for Data Collection Projects on our Self-Service platform. Come and try the updated flow yourself at https://toloka.ai/platform