Meramia's knowledge work already lives inside boundaries the client is held to. This role makes those boundaries measurable: privacy-enhancing techniques on the training and inference path, red-team tests for leakage, and tools the rest of the lab can run without a paper as a crutch.
You will turn a result from the literature into code that ships, and you will say when a privacy guarantee is real and when it is a slide.
What you will do
- Prototype and land privacy-enhancing algorithms — differential privacy, secure aggregation, federated learning — on the training and inference stacks we actually run.
- Red-team the models for membership inference, inversion, and training-data memorisation. Write the failure down before someone else finds it.
- Build the evaluation suites and diagnostic libraries ML engineers can use across a model's life, not a one-off notebook.
- Sit with research, platform, security, legal and product so a regulation or a security principle becomes a guardrail in code.
- Investigate the privacy–utility trade: capability, latency and cost against a guarantee you can state.
What you bring
- PyTorch or JAX, and research-grade Python you will test rather than demo.
- Differential privacy (including DP-SGD), secure multiparty computation, or federated learning — implemented, not only cited.
- The attack surface: extraction, membership inference, poisoning — and how you measured it.
- A paper you can turn into a well-tested module without losing the claim that made the paper worth reading.
- The ability to explain a mathematical privacy guarantee to an engineer and to a policy lead in the same week.
Useful, not required
- Peer-reviewed work or open-source in privacy, security, cryptography or machine learning — NeurIPS, ICLR, USENIX Security, IEEE S&P, or the equivalent venue.
- PETs on a distributed train or a high-throughput inference path, not only on a single node.
What to send
- A CV and the two profile links the form asks for.
- One privacy or leakage result you implemented — the threat, the metric, and the trade you accepted.
The first quarter
The first month is an audit of the privacy evaluations and training workflows we already run. The second is one automated evaluation or a differential-privacy module in the internal ML path. The third is an internal note on privacy–utility for the architectures we serve, with a recommendation that can be deployed rather than filed.