I spoke with Mikhail Baklanov, CEO and founder of Donkit, about a problem every enterprise AI team hits sooner or later, retrieval augmented generation that is accurate enough for production. The numbers he quotes are familiar. Companies spend heavily and still spend 18 months experimenting with indexes, chunking, embeddings, evaluators, and guardrails. Accuracy is often not good enough at the end.
Donkit’s pitch is simple. Give their “RAG ops” agent your goal and your data. One engineer sets it loose. It runs hundreds of experiments and returns a production ready configuration in about two days. The team pivoted to this idea in March. By August they had a closed alpha. A dozen enterprise teams are piloting it now.
Baklanov frames RAG like a library, with indexes, storage, and a librarian. Donkit is a library factory. Instead of hand tuning for months, the agent explores the space of options, evaluates on your data, and converges on what works. The promise is less art project, more systematized process for context engineering and memory management.
Their buyer is the head of AI, or a principal AI engineer with a mandate to ship internal assistants and agents. Donkit is not for small teams standing up a single chatbot. It is for accuracy sensitive use cases where error compounds across steps. As Baklanov puts it, if your RAG layer is 80 percent accurate and an agent queries it five times in a chain, the final step’s effective accuracy can collapse. That is why enterprises throw so much effort at squeezing out the last five to ten percent.
One pilot sits inside a large retailer’s HR call center. Fifty three specialists support three thousand employees across time zones. Routine questions can take thirty minutes in SAP. Donkit’s approach augments the specialist in real time, first through typed suggestions, then by listening to calls and surfacing answers on screen. It is a clear ROI case that lives or dies on reliable retrieval.
Baklanov has talked with more than eighty heads of AI. He sees the same adoption gap. People do not understand how AI works or how to use it. The wins show up in augmentation, not replacement. In software development he says the job is already shifting from hand coding to tasking and reviewing AI generated code. Speed of iteration trumps the debate over organic versus AI code.
Donkit has raised $470,000 in angel funding and is preparing an institutional seed. The site is donkit.ai. If you are a larger organization with a head of AI, running agents that depend on trustworthy retrieval, they want to talk.
RAG is where a lot of AI projects stall. Everyone has a demo. Few have a system that survives contact with real data and real users. If Donkit can consistently compress the tuning loop from 18 months to two days and do it with one engineer instead of a roomful, that changes the economics. It also sets a higher bar for what “production ready” means.
