Role Overview
Nightlet is a self-awareness company. We build emotional intelligence that finally understands your heart, body and stars: an app and an optional bracelet for people who want to understand themselves, not optimize themselves. We combine what people say, how they sound, what their body does and the sky they live under into one honest picture, and we turn that picture into calm.
As our Research Engineer, you build the models behind SOUL, our State Observation & Understanding Layer: speech, audio and multimodal models that turn words, voice and signals into an understanding of how someone actually is. You turn open questions into experiments, train and refine models, and build the software that makes training, inference and benchmarking reproducible.
This is a hands-on individual contributor role for someone who enjoys open-ended investigation as much as the code that makes research useful. You own the technical work from the first research question through experiments, analysis and validated model improvements, on models we develop in-house and models partners bring to us.
What you’ll do
- 01
Explore research questions in speech-language, audio and multimodal machine learning and design experiments that test how models behave and what they are capable of.
- 02
Train, fine-tune, validate and keep improving the models built at Nightlet as well as the models our partners bring to us.
- 03
Build dependable software for distributed training, inference and benchmarking, so experiments can be repeated and results compared.
- 04
Set up data workflows for collecting, storing, preparing and analysing the large datasets we train and evaluate on.
- 05
Define evaluation protocols and benchmarks that capture model quality, generalisation, robustness and progress against our research goals.
- 06
Dig into datasets, model outputs and failure cases, compare training approaches and let the evidence decide the next round of development.
- 07
Work with partner research and technical teams to define the questions, understand their models and data, and run training, validation, evaluation and further development together.
- 08
Write down experiment designs, results, limitations and technical decisions so colleagues and partner teams can reproduce the work and build on it.
What you bring
- 01
Two or more years of contributing to model training or fine-tuning on large text, audio, image or video datasets, with real depth in speech, audio or multimodal machine learning.
- 02
Strong Python and PyTorch, including developing, debugging and maintaining training and evaluation code.
- 03
Experience writing robust, maintainable code for research systems that live beyond a notebook or a first prototype.
- 04
Sound experimental judgement: forming hypotheses, setting useful baselines, designing evaluations and reading results with care.
- 05
Proof of research and model-building ability: publications, model releases, shipped systems or substantial open-source work.
- 06
Comfort iterating quickly on uncertain research directions and taking ownership of open-ended technical problems.
- 07
Clear written and spoken communication, including explaining methods, findings and limits to researchers and to partner engineering teams.
- 08
A PhD is not required. We care about demonstrated research ability and the quality of the models, experiments and systems you have helped build.
Bonus points
- 01
Training or fine-tuning transformer models.
- 02
Depth in text-to-speech, speech recognition, speech-to-speech, prosody, emotion recognition or audio understanding.
- 03
Post-training, preference learning or human evaluation of generative models.
- 04
Making distributed training, inference or large-scale data processing more efficient.
- 05
Collaborating with external research teams on model development or evaluation.
Sounds like you?
A CV, a link to something you have built and a few honest sentences are enough to start.
Apply now →How we work
We build Nightlet the way we hope people will live with it: understood, not optimized. That starts with how we work together. Nobody here keeps score of your hours, so if you need a run in the middle of the day, you take it. Sport is unlimited and everyone receives a spiritual allowance for whatever restores them, from meditation to therapy.
We test what we build ourselves and we tell each other honestly when something doesn’t feel right. For you that means your models are used by the team every night, not only benchmarked.
Closing note
If you have ever wondered what a voice reveals that words don’t, you already know why this role matters. We’re looking forward to hearing from you.