Scientifique de données / Data Scientist

Ondova•Montreal, QC

9d•Onsite

About The Position

Ondova is a Montreal deep-tech startup, spun out of the TandemLaunch incubator, building headphone technology: we reuse an ordinary headphone or earbud driver as both the audio speaker and a biosignal sensor, with no added sensors. That turns everyday headphones into heart-rate, wear-detection, user-ID, and gesture sensors. We are pre-commercial, patent-backed, and building the reference hardware and software that partners and OEMs will license. You will work directly with our technical lead on the science at the core of the product: the algorithms that turn raw speaker-sensed signals into biometrics, heart rate and related vitals, user identification, wear detection, and gesture recognition. The data acquisition tooling already exists; your job is the data pipeline: data acquisition, extracting physiological signal from noise, building and evaluating models, and testing hypotheses, how well each feature holds up under motion, playback, and real-world variation. This is a hands-on research-engineering role for someone who wants their models to end up in a shipping product rather than a paper. You will own the quality of our results: the evaluation harnesses, the benchmarks, and the discipline.

Requirements

Master’s or PhD in engineering, computer science, physics, applied math, or a related field; or a Bachelor’s with equivalent research experience; and roughly 3-8+ years of applied work; we will level the role to the candidate.
Strong Python scientific stack (NumPy/SciPy/pandas; scikit-learn and/or a deep-learning framework) and solid software habits: tests, Git, code review.
Real signal-processing fundamentals: filtering, FFT and spectral analysis, sampling, so you can reason about a time series before you model it.
Statistical and evaluation literacy: you know when correlation misleads, why Bland-Altman exists, and how to design an experiment that could prove you wrong.
Comfortable with messy real-world data and with stating assumptions clearly.
A clear communicator who can explain a result, and its limitations, to both researchers and business stakeholders.

Nice To Haves

Experience with biosignals (ECG, PPG, audio).
Time-series deep learning; self-supervised or small-data methods.
Awareness of real-time and on-device constraints when designing algorithms (compute, latency, memory).

Responsibilities

Explore and understand speaker-sensed biosignal datasets; form and test hypotheses about what the signal contains and what corrupts it.
Design, implement, and evaluate signal-processing and machine-learning methods for feature extraction, heartbeat recovery, wearer identification, gesture classification, and for interference rejection.
Own results QA: build evaluation harnesses and regression benchmarks (e.g., MAE, MAPE, correlation, Bland-Altman, availability), track them as methods evolve, and report what works and what doesn’t with equal clarity.
Run robustness studies: motion artifacts, playback conditions, fit variation, unit-to-unit spread, quantify where each feature degrades and why.
Document methods and results so that every claim we make to partners is traceable to an experiment.