Charly answers the phone for more than 100 German municipalities — around the clock, in over 30 languages, hosted in the EU. Behind it is a small team in Berlin that ships fast: around 150 pull requests land in our repositories every month, and a good idea often talks to real citizens the same week. On the phone, the speech layer decides everything: does Charly catch a street name spoken in dialect, does the answer sound natural, does Charly avoid talking over people? We are now building this layer ourselves — with open models, on our own GPUs in Germany. That's where you come in.
Your Role Select open STT and TTS models, fine-tune them on our domain with PyTorch and measure them properly against the baseline
Serve models in production on our own GPUs in the EU: streaming, batching, quantisation, latency budgets and cost per call
Build the voice agent's orchestration layer: turn-taking, barge-in, tool calls and the hand-over to our LLM agent (Python, FastAPI, pydantic-ai)
Evals that gate releases: word error rate on anonymised calls, pronunciation and naturalness tests, end-to-end conversation evals — wired into CI
Monitor quality in production and close the loop: from a failed call to a test case and to better training data
Your Profile We welcome experienced candidates with several years of work on ML systems just as much as career starters who have mostly worked in notebooks so far
Good PyTorch skills, ideally with hands-on experience fine-tuning speech or audio models
Ideally experience with STT and/or TTS, e.g. Whisper-family models or modern neural TTS
Experience with serving (GPU inference, streaming, quantisation, latency under load) is a plus — otherwise you will learn it with us
A strong quality mindset: you don't ship without an eval, you distrust single examples, and you build benchmarks that predict real-world quality
You use AI coding assistants every day and move fast without putting production at risk
Nice-to-have: real-time voice (telephony/SIP, VAD, barge-in), privacy-preserving ML, German language skills
What We Offer Real ownership: your models talk to citizens on live phone lines within your first month
Our own GPU infrastructure in the EU and room to experiment
A beautiful and central office right at Rosenthaler Platz
28 days of vacation per year
Our Application Process 01
Application to jobs@zaitgeist.com 02
Short phone screen 03
On-site interview Python
FastAPI
Supabase
Frontend
TypeScript
Next.js
Tailwind
Cloud
EU hosting
Docker
Coolify
Logfire
AI / ML
pydantic-ai
PyTorch
vLLM
Ready for the next step? We look forward to meeting you. Send us your application with your
CV and relevant projects or code samples.
To filter out AI-generated applications, please start your cover letter or
#J-18808-Ljbffr
