Machine Learning Resident – Client: OpenCycle (12 month term)

Amii (Alberta Machine Intelligence Institute); Edmonton; On-site; CAD 42,000 - 66,000; Full time

Job summary

Amii (Alberta Machine Intelligence Institute) is seeking a resident to work with OpenCycle on acoustic intelligence for industrial sites. You will contribute to ML research, collaborate with a cross-functional team, and receive mentorship from renowned scientists during a 12-month paid residency.

Successful candidates will engage with environmental noise compliance challenges, model deployment on edge devices, and knowledge transfer activities, with potential for hire by the client OpenCycle

Qualifications

• Solid background in machine learning, specifically deep learning for audio and acoustic signal processing - sound event detection.
• Experience with audio representation learning and evaluation that holds up under real-world domain shift.

Responsibilities

• Design, implement, optimize and evaluate models for far-field acoustic sensing: sound event detection and open-set source identification.
• Prepare and preprocess high-quality audio datasets for training, including feature front-end design (log-Mel, PCEN, normalization).

Skills

Machine learning

Deep learning (audio)

Acoustic signal processing

Sound event detection

Representation learning

Job description

\"If you are interested in the application of machine learning to real-world audio for monitoring the noise industry puts into people's lives, this is the right opportunity for you. Be a part of the team of research and machine learning scientists building acoustic intelligence for industrial sites from the ground up and get mentored by some of the best minds in AI during the process.\"

• Kunwar Saaim, Machine Learning Scientist, Advanced Technology

About The Role

This is a paid residency that will be undertaken over a 12-month period with the potential to be hired by our client, OpenCycle, afterwards (note: at the discretion of the client). The Resident will report to an Amii Scientist and regularly consult with the client team to share insights and engage in knowledge transfer activities. Successful candidates will be members of a cross-functional project team with backgrounds in ML research, project management, software engineering, and new product development. This is a rare opportunity to be mentored by world-class scientists and to develop something truly impactful.

About The Client

OpenCycle is a Calgary-based acoustic compliance and site-management platform for the energy, municipal, and heavy industries. The company grew out of four decades of professional acoustics consulting: its founders and senior staff have spent their careers on noise impact assessments, complaint investigations, mitigation design and regulatory negotiation across Alberta, British Columbia, Saskatchewan and Manitoba. OpenCycle exists to turn that hard-won expertise into software, so that noise compliance - which has traditionally meant months of specialist fieldwork and manual reporting - can be predicted, documented and cleared in days.

Today the platform combines regulatory modelling, site and asset management and a growing fleet of in-house-designed acoustic monitoring hardware deployed at customer sites across Alberta and BC. Machine learning is not a side project here - it is the core of the company’s next generation of product, developed by an in-house engineering team working directly alongside practising acousticians. That proximity is the point: models are specified, labelled, sanity-checked, and ultimately signed off by domain experts inside the same organization, and there is a short, direct line from a research result to a sensor running in a field.

The company’s mission is to reduce the impact industrial emissions have on people’s lives. Noise is where OpenCycle starts, because it is the emission that most directly affects the communities living next to energy and infrastructure development - and because it is the problem this team knows better than anyone.

About The Project

Environmental noise compliance today answers one question well: how loud was it? A sound level meter returns a number. What it cannot say is what made the noise. Attribution - deciding which of the several sources on and around a site is responsible for the level measured at a home or a receptor - is still done by an acoustician listening to recordings by hand. It is the most expensive and least scalable step in the entire compliance workflow, and it is the step this project automates.

The technical problem

The system has to answer four questions from a single learned representation: is a sound source present, what kind of source is it, which specific physical unit is it and is that unit operating normally? It has to do this in the open air, where several sources overlap continuously and the interesting one is rarely the loudest. And it has to do it on a battery-powered outdoor node with microcontroller-class compute, reporting over a long-range radio link whose payload is measured in tens of bytes. Sending the audio to a large cloud model is not an option, so the central research question is how much of this capability survives compression to an edge budget.

What already exists

This is not a greenfield exercise. There is a deployed fleet of sound-level-meter nodes running a separately certifiable IEC 61672 measurement chain and validated against reference instruments in the field; a working detection-plus-embedding architecture with quantized, radio-sized payloads; an on-device inference path verified stage-by-stage against the research reference; and a frozen benchmark protocol with cross-site and cross-device evaluation tiers, built deliberately to expose the failure modes this team has already been burned by. There is also a substantial written record of experiments - including the ones that failed, which are often the more useful half.

The open problems

Four, and the Resident would help shape which ones to attack. First, channel invariance: our identity embeddings are currently key on the recording channel rather than the source, and six remediation strategies plus the data-scale hypothesis have been eliminated under controlled comparison, pointing at a label-ontology root cause. Second, tracking and temporal accumulation: fusing observations of the same source over time is the single largest measured improvement we have, turning a hard single-look problem into a tractable one - making the tracker first-class and validating it on real polyphonic field recordings is the dominant unknown before deployment. Third, condition detection from scarce labels: whether the operating state can be represented separately from identity when almost no per-unit multi-state data exists publicly. Fourth, compression: distilling the cross-site stability of a large pre-trained audio backbone into something that fits the sensor’s compute and bandwidth budget.

What the year looks like

Real deployed hardware and real field data rather than a benchmark exercise; a measurement culture that treats negative results as results and freezes evaluation protocols so experiments months apart stay comparable; scope to publish; and a direct path from a research result to a device in a field that changes how a regulator makes a decision affecting a community. The Resident will work with OpenCycle’s engineering team and its acousticians, with field access to the sensor fleet and to the domain experts who produce the ground truth.

Required Skills / Expertise

Are you passionate about building great solutions? You’ll be presented with opportunities to both personally and professionally develop as you build your career. We’re looking for a talented and enthusiastic individual with a solid background in machine learning, specifically deep learning for audio and acoustic signal processing - sound event detection, audio representation learning, and evaluation that holds up under real-world domain shift.

Key Responsibilities

• Design, implement, optimize and evaluate models for far-field acoustic sensing tasks - sound event detection, open-set source identification and operating-condition / anomaly detection - from a shared learned audio representation.


• Prepare, curate and preprocess high-quality audio datasets for training or fine-tuning, and validating models, including feature front-end design (log-Mel, PCEN, per-channel and frequency-wise normalization), audio