Nightlet isaself-awareness company. Webuild emotional intelligence that finally understands yourheart, body andstars: anapp andanoptional bracelet forpeople who want tounderstand themselves, notoptimize themselves. Wecombine what people say, how they sound, what their body does andthesky they live under intoone honest picture, andweturn that pictureintocalm.
As ourResearch Engineer, youbuild themodels behind SOUL, ourState Observation &Understanding Layer: speech, audio andmultimodal models that turn words, voice andsignals intoanunderstanding ofhow someone actually is. Youturn open questions intoexperiments, train andrefine models, andbuild thesoftware that makes training, inference andbenchmarkingreproducible.
This isahands-on individual contributor role forsomeone who enjoys open-ended investigation as much as thecode that makes research useful. Youown thetechnical work fromthefirst research question through experiments, analysis andvalidated model improvements, onmodels wedevelop in-house andmodels partners bringtous.
What you’ll do
- 01
Explore research questions inspeech-language, audio andmultimodal machine learning anddesign experiments that test how models behave andwhat they are capableof.
- 02
Train, fine-tune, validate andkeep improving themodels built atNightlet as well as themodels ourpartners bringtous.
- 03
Build dependable software fordistributed training, inference andbenchmarking, so experiments can be repeated andresultscompared.
- 04
Set up data workflows forcollecting, storing, preparing andanalyzing thelarge datasets wetrain andevaluateon.
- 05
Define evaluation protocols andbenchmarks that capture model quality, generalization, robustness andprogress against ourresearchgoals.
- 06
Dig intodatasets, model outputs andfailure cases, compare training approaches andlet theevidence decide thenext roundofdevelopment.
- 07
Work withpartner research andtechnical teams todefine thequestions, understand their models anddata, andrun training, validation, evaluation andfurther developmenttogether.
- 08
Write down experiment designs, results, limitations andtechnical decisions so colleagues andpartner teams can reproduce thework andbuildonit.
What you bring
- 01
Two ormore years ofcontributing tomodel training orfine-tuning onlarge text, audio, image orvideo datasets, withreal depth inspeech, audio ormultimodal machinelearning.
- 02
Strong Python andPyTorch, including developing, debugging andmaintaining training andevaluationcode.
- 03
Experience writing robust, maintainable code forresearch systems that live beyond anotebook orafirstprototype.
- 04
Sound experimental judgment: forming hypotheses, setting useful baselines, designing evaluations andreading resultswithcare.
- 05
Proof ofresearch andmodel-building ability: publications, model releases, shipped systems orsubstantial open-sourcework.
- 06
Comfort iterating quickly onuncertain research directions andtaking ownership ofopen-ended technicalproblems.
- 07
Clear written andspoken communication, including explaining methods, findings andlimits toresearchers andtopartner engineeringteams.
- 08
APhD isnotrequired. Wecare about demonstrated research ability andthequality ofthemodels, experiments andsystems youhave helpedbuild.
- 01
Training orfine-tuning transformermodels.
- 02
Depth intext-to-speech, speech recognition, speech-to-speech, prosody, emotion recognition oraudiounderstanding.
- 03
Post-training, preference learning orhuman evaluation ofgenerativemodels.
- 04
Making distributed training, inference orlarge-scale data processing moreefficient.
- 05
Collaborating withexternal research teams onmodel developmentorevaluation.
Sounds like you?
ACV, alink tosomething youhave built andafew honest sentences are enoughtostart.
Webuild Nightlet theway wehope people will live withit: understood, notoptimized. That starts withhow wework together. Nobody here keeps score ofyourhours, so if youneed awalk inthemiddle oftheday, youtake it. Sport isunlimited andeveryone receives aspiritual allowance forwhatever restores them, frommeditationtotherapy.
Wetest what webuild ourselves andwetell each other honestly when something doesn’t feel right. Foryouthat means yourmodels are used bytheteam every night, notonlybenchmarked.
Closing note
If youhave ever wondered what avoice reveals that words don’t, youalready know why this role matters. We’re looking forward tohearingfromyou.
#J-18808-Ljbffr