Senior Applied Voice AI Research Engineer - Poland

Voicegain; Polska; Remote; PLN 354,000 - 512,000; Full time

Job summary

Voicegain builds speech-to-text and voice AI infrastructure. We’re hiring one researcher who ships, at the intersection of LLMs and speech recognition, solving failures across multilingual support and real-time processing.

You’ll work directly with the founders, formulate problems, prototype, measure, and push production-ready solutions into live traffic, balancing research depth with pragmatic deployment.

Qualifications

• PhD in ML / NLP / speech or depth of research experience is a strong plus.
• Hands-on with LLMs beyond prompting, fine-tuning with LoRA/QLoRA or similar.
• Deep experience in ASR, NLP, or related speech technologies.
• Strong Python skills and ability to ship production code in large Java/C++ codebases.

Responsibilities

• Tackle ML and speech-related research problems in ASR and NLP.
• Prototype, measure, and deploy production-grade solutions.
• Collaborate directly with founders to move experiments to production quickly.
• Balance research depth with practical deployment constraints.

Skills

ML foundations

Fine-tuning models

Python proficiency

Production software

Kubernetes

NVIDIA stack familiarity

Education

PhD in ML / NLP / speech

Tools

LoRA/QLoRA

MRCP

Job description

Voicegain Fully remote Compensation competitive for your market

Voicegain builds speech-to-text and voice AI infrastructure. Our ASR runs inside production telephony systems — IVRs, contact centers, voice bots — where the audio is noisy, callers don’t follow the script, and “works on the benchmark” is not good enough.

We’re hiring one person: a researcher who ships.

The problems

You’ll work at the intersection of LLMs and speech recognition, where many of the interesting failures live. A few examples of what “solved” looks like here:

• A redaction system that finds a credit card number in a transcript where the caller paused, corrected themselves, numbers were partially repeated by the agent, then corrected again.
• AI grafted onto legacy ASR platforms: sometimes there’s no clean slate, and your AI has to live inside an existing stack — behind an MRCP interface, for example — without breaking what already works.
• All of it working in Spanish, French, Polish, and other languages — not just English.
• An ASR that takes incomplete or imperfect GRXML grammars and interprets callers’ utterances the way a human would. The grammar says “pay bill”; the caller says “I’d like to pay off what I owe”; the right semantic tag comes back anyway.

There is no established playbook for much of this. You’ll help write it.

How we work

You get a problem, not a spec. You formulate it, read the papers, analyze deployment constraints, pick candidate approaches, prototype, measure on real data, analyze failures — then build the production version and watch how it behaves on live traffic.

You work directly with the founders and have a short path from a successful experiment to a production implementation.

You

• Deep ML foundations. You understand how transformers, embeddings, and fine-tuning actually work — well enough to adapt or combine techniques when nothing off-the-shelf fits. A PhD in ML / NLP / speech, or research experience of that depth, is a strong plus.
• Hands-on with LLMs beyond prompting. You’ve fine-tuned models (LoRA/QLoRA or similar), evaluated them rigorously, and know when a small classifier beats an LLM.
• Serious depth in speech or NLP. ASR, spoken-language understanding, semantic similarity, information extraction — you’ve done real work in at least one.
• A real engineer. Excellent Python. You can navigate a large Java or C++ codebase and ship production code, not just notebooks.
• Comfortable with production ML infrastructure. GPU inference, Kubernetes, model serving. Experience with the NVIDIA stack (NeMo, Riva, Triton, TensorRT) is a plus.
• A highly effective user of modern AI development tools — with the judgment to validate what they produce.
• Independent. You make sound technical decisions with minimal supervision and can explain and defend them.

We don’t care whether you’ve worked on exactly these applications before. We care whether you have the depth, breadth, and curiosity to attack problems whose answers aren’t known yet.