Interview
By
Tendem Team
"Can You Trust the Result?" Meet the Engineer Behind Tendem's Orchestrator
Slava built the brain of Tendem – the system that decides whether your task needs AI, a human, or both, and tells you the price before you commit. We asked him how it works.
Slava is an ML team lead at Toloka and one of the people behind Tendem’s orchestrator. This conversation is edited lightly for length and clarity.
When a tool finishes a job, it hands you the output and says: here are the results, go ahead and use them. The catch is the question that follows. Can you trust them? Are they really what you expected? For most AI work today, answering that question is left to you.
Most of the industry is racing to remove the human from the loop. Tendem is doing the opposite – deliberately putting a vetted human expert inside the workflow, and building a system to decide when that human is needed. That system is the orchestrator, and Slava is one of the people who built it.
Tendem 101 – what is it, in the simplest terms?
It is a service that gets tasks done – that pretty much sums it up completely. It is different from the regular tools you might use, because while those help a lot, they do not actually guarantee they can solve any task. There are too many peculiarities across domains like design or mathematics. And it is different from freelance platforms, because you do not have to manage freelancers, discuss the price, argue over scope, and organize the workflow. You just give Tendem a task and it delivers a high-quality result.
Lift the curtain a little. What is inside Tendem that lets it solve almost any task?
It is not something completely new that we invented. We saw the advantages of using AI, and we also understand the advantages of asking humans and using their input and their work. We decided to mix the two and get the highest quality at the best possible speed. The only problem is that this requires orchestration – a management mechanism that oversees the whole process and produces the result using AI and humans as tools.
“Human in the loop is not something Tendem invented. It is what we are basing the technology on.”
So something inside has to decide: AI, a human, or both. How does that work?
That is the crucial question behind the orchestrator, and it is an evolving research area. The mechanism has to understand which tasks an AI can complete successfully with high probability – say 95% – and which it cannot. If a task only has, say, a 67% chance of success, the orchestrator has to recognize that and do something about it. For example, it can run a QA cycle after getting the result and check whether it actually worked.
For readers who don’t know the term – what is QA filtering?
When something gets done, it is great that a tool tells you “I did the job.” But can you trust the result? You used to assess it manually – this one is good, this one is bad, go back to the freelancer and say we did not agree on this, please redo it. The same happens with AI: you give it a prompt, it does something, you say “no, that is not what I wanted,” and you loop like that until you finally get a good result, or you never do because the model simply cannot do what you are asking.
The answer is a mechanism that assesses results for you. One way is automatic – an LLM checks the quality of a result. It is cheap and easy, but it is not 100% correct either, so you end up having to QA the QA model, and that loop continues forever. So we also have human QA, where real people assess whether results are good. Combining the two gives us the best quality when we need it, and the best speed when we need speed instead.
A lot of companies treat the human as a crutch to eliminate. Tendem goes the other way. Why?
AI and humans each have their own capabilities. AI is getting better every month, but there are still domains and problems out of scope for current models – and it is not only about domains. It is about the ideas behind a task and the nuances of context a model should have. AI does some things really well and it is a great tool. But it has drawbacks, and with a human you get oversight and taste.
Take writing an article for your blog. You can draft it with AI easily, but most people can spot AI-generated text – even if you strip out the tells, like swapping the long dashes for hyphens, there are still words and phrasings AI overuses. Readers notice, and I am fairly sure search engines can tell too, and may penalize a site that keeps posting AI slop. So you need a human touch: someone who reads it and says this part reads like AI, let us rewrite here and here.
Context nuance is the harder problem, and it grows with the stakes. Until recently, ask a model “I’m 50 meters from the car wash, should I drive or walk?” and it would tell you to walk because 50 meters is short – missing the entire point. Now scale that up: a good landing page needs design, copywriting, optimized code, and product sense – calls to action, an FAQ, the right sections. That is why we built a mechanism that breaks a complicated task into parts, each executed in the most optimized way by AI or by humans.
If the orchestrator does the heavy lifting, what is left for the user to decide?
We leave the user the most important interaction: give your input and settle the scope. That is it – what you want to achieve and what result you want. It also solves a problem freelance platforms have, where you have a task but then also have to choose a freelancer, work out the price, discuss everything, manage the time, and QA the result at the end because they might just send you something AI-generated. We take all of that off the client. We help you decide the scope, then we do the magic under the hood – breaking it down within a budget into AI execution, human execution, and QA sessions – and guarantee a high-quality result.
Tendem charges per task and tells you the price up front. What’s the thinking there?
It is an interesting story. With AI tools you are used to paying as the model runs, and you can quickly hit token limits and sit waiting hours for them to refresh – or watch an agent get stuck in a loop, burning tokens. With freelancers, you negotiate: you want to pay less, they want more. We got feedback from early users that it was not clear how much a task would cost, because we used the same post-paid model as AI tools – we told people the cost only after running the task.
So we reversed it. We predict the price up front. We built an estimator – a fairly classic machine-learning application with several models – that reflects current market prices and covers our costs for AI execution and for paying our human experts. It runs on your task scope, so the same scope gives the same price. Want more records in a lead list? You pay more. A single article? Less.
Rough ranges Slava shared for what different tasks tend to cost:
Task type | Typical price | Turnaround |
Quick web research (a few sources) | A few dollars to ~$20 | Fast |
Article or blog post | Around $40 – $50 | Same day |
Logo or single design asset | Around $40 – $50 | Same day |
Complex, multi-step task | From $100 | One to two days |
Ranges are indicative and depend entirely on scope; you see and approve the price before any work starts.
Compared with hiring a freelancer, it is faster and often cheaper – no platform fee, no time lost finding, briefing, and managing someone, and AI speeds up the work. We keep a margin, but because we understand when to use humans and when to use AI, we can deliver higher quality than AI tools alone and, in many cases, higher quality than freelance platforms. We published a white paper earlier this year comparing this; some numbers have moved since, but the idea holds.
What’s the hardest part of building this, as the engineer behind it?
Mixing AI and humans is new-ish for the industry, though not for the research community. There are different ways to do it – AI and a human working together on the same task, or one after another on smaller jobs. It is not simple. We found, as others have, that on some tasks the combination actually slows things down while keeping quality high. Working together means sharing context: the AI has to explain what it did and why one approach worked and another did not, and the human explains back. That communication takes time. So some tasks are simply better given to a single actor, and the orchestrator’s job is to route the data flow well.
Where does this go in two or three years?
Models are getting much better at orchestration – breaking tasks into subtasks and understanding how the pieces interact. That will keep improving. AI capabilities will grow in the domains where AI already does well, but they will not cover everything. There will always be domains where a human is a necessity – you simply cannot solve the task without one. So expect a shift in which domains lean on humans, but the interaction between AI and humans is here to stay for quite a while. It is also good to see more players entering the market with their own versions of how this combination might work.
What’s a task you’ve run through Tendem yourself that you loved the result of?
One favorite: I wanted high-resolution images to print as posters for my room and could not find them online. Upscaling tools did not work well. I gave the task to Tendem, and our experts working with AI actually delivered what I wanted – I was amazed, and I kept using it. The most recent one was finding places to pay my water bills. Tendem nailed it: a clear list, each point on a map near my house, with who to talk to and where to go. Now, whenever I have something I do not have the time or resources for and I do not want to babysit an AI tool through a QA loop, I just give it to Tendem.
It is a glimpse of where the product is heading: something closer to a virtual co-worker – an always-available teammate you can reach right inside Claude or ChatGPT through the new MCP connector, that picks up wherever you left off.
Bring the task you’d normally have to babysit.
Give Tendem your scope in plain language. The orchestrator decides what needs AI, what needs a human, and what needs both – and shows you the price before anything runs. No freelancer to manage, no QA loop, no surprise invoice.
