We use PostHog and Clarity to improve the site, and Meta Pixel to measure page views and ad effectiveness. Usage data may include page addresses and browser identifiers. You can opt out of collection at any time. Details in the Privacy Policy
More fit checks, tailored CVs, and complete prep for every role.
Job Description
What you'll do
You take a novel research proposal that works in a small-scale proof of concept and find out whether it holds up for real — across model sizes, tasks, datasets, and hardware constraints. When it breaks, you figure out why and build what it takes to fix it, working side by side with the researcher who proposed it.
Scale a proof of concept into a model zoo: a broad set of model sizes, tasks, and datasets that stress-tests the architecture well beyond the original demonstration.
Diagnose where it breaks as it scales — and it can break in any of four places: the math (does the theory hold at scale?), the architecture (do small-scale design choices still hold?), the data (does it generalize beyond what it was first validated on?), and the system (does it survive real compute, memory, and latency constraints?). You move between all four.
Work directly with the researcher who built the proof of concept to close the gap between what worked in a controlled setting and what runs efficiently on Axonica's silicon.
Build the evaluation infrastructure that makes this repeatable rather than one-off: benchmark harnesses, ablation tooling, and reference implementations that feed directly into the SDK, compiler, and quantization paths so a validated idea actually runs on Axonica's hardware.
Earn the room to lead your own research directions. As your judgment on the architecture proves out, you will propose and drive new lines of work, not just scale up someone else's idea.
Requirements
3 to 5 years of hands-on, real-world experience in AI/ML engineering or applied AI/ML research. We are looking for people who have already shipped or stress-tested model work against real constraints, not a graduate or first role profile.
Comfort diagnosing failure across math, architecture, data, and systems, rather than only one of those layers. Most candidates are strong in one dimension; we need people who can reason across all four.
Direct experience with model scaling work: taking something that worked at small scale and figuring out why it does or does not hold at larger scale.
Working knowledge of several efficient-inference techniques — for example quantization, low-bit pipelines, KV-cache compression, speculative decoding, or kernel- and memory-level optimization — and familiarity with at least one non-Transformer model family (e.g. SSMs, diffusion models), relevant to translating research ideas into something that runs efficiently on real hardware.
Comfortable working without a fixed spec. You will often be taking something that is partially proven and figuring out, with the Science Researcher who built it, what it takes to make it actually work.
Strong Python fundamentals (C++ a plus), and systems fluency to work close to the hardware boundary (compiler, runtime, hardware abstraction layer) when the architecture demands it.