Conversational AI
A voice assistant answering a retail bank's phone line
Part of the program that built Oleg, Tinkoff Bank's Russian-language voice assistant for retail customers.
- Employer
- Tinkoff Bank
- Period
- 2019-2020
- Stack
- Speech recognition, Text to speech, Intent classification, Dialogue management, Telephony
The problem
A retail bank at national scale takes an enormous volume of calls, and most of them come down to the same few questions. The standing answer was a deeper menu tree: press one, press two, hold, reach a person, explain yourself again. Customers do not want a menu, they want the answer. The bank needed something that could listen to a sentence in Russian and act on it.
What I worked on
Oleg was a production spoken dialogue system for the bank's retail customers, reachable by phone and inside the app: Russian speech recognition over a phone channel, classification of what the caller actually wants, dialogue state that survives interruption and correction, and a handover rule for the moment the assistant should stop guessing. I worked inside that program rather than owning it, alongside the speech, product and operations teams that each held their own piece. What is easy to underestimate from outside is how much of the work sits away from the model: reviewing transcripts, building the taxonomy of real requests, and arguing about what the assistant is allowed to say when it is not sure.
Result
Shipping conversational AI to live retail customers teaches one thing hard: the model is the cheap half, and the expensive half is what the system does when it is wrong in front of a customer. The confidence thresholds, escalation paths and human review queues in everything I have built since come from there, along with the habit of counting how many real people actually got an answer.