Bitmanager.ai — AI startup in one of the fastest-growing categories of conversational AI agents.
We are building a platform for voice and text AI agents in customer communications: sales, support, consultations, and debt collection.
Our agents work with data, access internal company systems, invoke tools, and conduct dialogues within defined business scenarios.
We are developing multilingual voice solutions and training models ourselves to expand the number of supported languages and improve product quality.
We are now looking for a strong voice technologies engineer who is responsible for results in real calls, not for metrics on an offline dataset. The specialist will take on one of the voice directions — speech recognition or synthesis — develop it for new languages, and be responsible for the operation of these services in production.
The main result of the work: an agent that hears the interlocutor, responds quickly, and sounds natural — in all supported languages.
What you will be doing
- Take responsibility for one of the voice directions — streaming speech recognition (ASR) or synthesis (TTS) — and develop it, including for languages with scarce data and ready-made solutions;
- Be responsible for latency: measure and reduce the time from the end of the interlocutor's utterance to the first sound of the agent's response;
- Develop dialogue naturalness: so that the agent understands in time when the interlocutor has finished their thought, calmly reacts to interruptions and pauses, and behaves like a human, not a robot, in a live conversation;
- Work with telephony specifics: narrow-band sound, streaming, echo, packet loss, analyzing recordings of real calls;
- Retrain and adapt speech models for our tasks and new languages, together with the data preparation team and language experts;
- Build a quality assessment system: offline recognition and synthesis metrics, end-to-end runs on real calls, regression tests before deployment;
- Build processes for deploying and operating speech services on GPUs — assembly, deployment, monitoring, incident analysis — and control their quality. You don't have to do everything manually, but the responsibility for ensuring these processes work lies with you;
- Trace call quality complaints to their specific cause: model, data, audio path, or infrastructure;
- Participate in the deployment of the voice part in the customer's infrastructure with the infrastructure team;
- Make technical decisions for the voice direction, review code, and help other engineers on the team grow.
What is important to us
- Proficient Python and experience developing production services, not just research scripts;
- Practical experience with speech technologies (ASR and/or TTS), including experience integrating models into production and their subsequent operation. Research experience only or model training only is not enough;
- Understanding of audio streaming processing: buffering, working with chunks, sources of latency, real-time constraints;
- Experience with PyTorch and running models on GPUs, understanding what inference time consists of;
- Ability to measure quality: select metrics, collect a test set, distinguish real improvement from noise;
- Operational experience: Docker, logs, monitoring, debugging problems that only reproduce in production;
- Ability to independently lead a direction, make decisions with incomplete data, and be responsible for the result;
- Ability to work with the product team, linguists, and technical specialists of the customer and clearly explain your decisions;
- Good Russian and English language skills: reading documentation and articles, communicating with the team.
Experience specifically with voice agents is not required. Engineering maturity, experience with sound, and the ability to quickly understand someone else's pipeline are more important.
Bonus points
- Experience with telephony: SIP, RTP, codecs, echo cancellation, working with 8 kHz sound;
- Experience building real-time voice agents or dialogue systems;
- Experience with VAD, detecting the end of an utterance, and handling interruptions;
- Experience with TTS retraining (SFT, LoRA) and working with voice cloning from a reference;
- Experience with multilingual speech models and languages with scarce data;
- Experience with inference optimization: batching, quantization, ONNX, TensorRT, vLLM;
- Experience with on-premise deployment of ML services in the infrastructure of large customers;
- Experience with LLMs as part of a voice pipeline.
If you are interested not just in improving metrics on an offline dataset, but in being responsible for how the agent hears, responds, and sounds in a real call — we are waiting for your application. We will be happy to meet you and discuss which voice direction you can take responsibility for.
