Senior ML Engineer

What matters most

• Turn verification needs into measurable objectives

  The role calls for defining positive classes, decision thresholds and leakage-free validation for problems such as deepfake detection in KYC. Evaluation must reflect asymmetric error costs, changing class prevalence and performance at very low false-positive rates.

• Build and own models end to end

  This is a hands-on model-building role, from data strategy and PyTorch training through deployment, monitoring and retraining. The posting emphasizes changing and explaining models, rather than integrating hosted APIs.

• Optimize inference with measured trade-offs

  At a stated peak of 2,000 requests per second, latency, throughput and cost-to-serve matter alongside accuracy. The role includes quantization, pruning and architecture changes, with production canaries before suitable workloads move from GPU to quantized CPU inference.

• Keep production models reliable

  The engineer is accountable for one or more model families in 24x7 production systems, including monitoring, drift detection and retraining under adversarial pressure.

• Raise the standard of model work

  The posting calls for reproducing research under production constraints and mentoring 3 to 6 junior engineers in mathematical reasoning, experimental design, reproducibility and model review.

Likely challenges to think through

• Choosing a threshold that holds up in live traffic

  Likely preparation challenge, not a confirmed problem: a strong aggregate score may conceal poor detection at the stated FPR of 1e-4, while prevalence shifts change precision. Prepare to explain how low-FPR results, calibration, leakage-free splits and error costs would guide an operating point.

• Reducing serving cost without losing critical accuracy

  Likely preparation challenge: quantization or a smaller architecture may save compute but alter predictions or latency under production load. Prepare a way to compare accuracy, throughput and cost, then specify what a production canary would need to verify before a GPU-to-CPU move.

• Maintaining performance as inputs and attacks change

  Likely preparation challenge: drift or new document and face attacks can make an offline result unreliable. Prepare to discuss cross-source or cross-generator testing, what to monitor after deployment and when retraining would be warranted.

Topics to prepare

• Statistical decision-making at low false-positive rates

  The posting emphasizes prevalence, asymmetric error costs, calibration and TPR at FPR = 1e-4. Revise by calculating how precision changes as deepfake prevalence falls, then explain why overall AUC may not identify the best operating threshold.

• Optimization and representation fundamentals

  The role calls for first-principles reasoning about loss functions, convergence and embedding geometry. Revise by deriving a classification or metric-learning loss, tracing its gradients, and explaining one cause of training divergence and a corrective experiment.

• Leakage and distribution shift

  The posting specifies leakage-free validation and cross-generator or cross-source tests. Revise by designing splits for a face or document model that keep related identities or sources apart, then explain what a failure on a new source would mean.

• Vision and NLP model choices

  Depth in either computer vision or NLP, with working knowledge of the other, is required; the model families include face detection, OCR and NER. Revise by comparing architectures and losses for one area of depth, then outlining how evaluation differs for one task in the other area.

• Inference numerics and cost tradeoffs

  The posting calls out INT8 and BF16 quantization, pruning, and latency, throughput and cost-to-serve. Revise by explaining how reduced precision affects a layer and its outputs, then specify which accuracy and serving measurements would justify a CPU-versus-GPU decision.

• Production model lifecycle and review

  The role includes monitoring, drift detection, retraining, 24x7 operation and mentoring junior engineers. Revise by sketching a model-review checklist covering reproducible experiments, production health signals, drift investigations and retraining decisions.

Skills to revise

• Implement and debug PyTorch training

  The posting explicitly requires training from scratch and fine-tuning, including training loops, pipelines, augmentation and loss design. As practice, write a small PyTorch training loop, deliberately introduce a convergence problem, diagnose it with logged gradients and losses, and fine-tune a pretrained network.

• Build an operating-point evaluation

  As practice for the stated evaluation work, compute confusion matrices, ROC and PR curves, calibration error, and TPR at a fixed low FPR on a held-out set. Check how a threshold performs under changed prevalence and document the limits of the estimate.

• Benchmark a quantized serving candidate

  The posting requires INT8 and BF16 experience and hands-on work with at least one of OpenVINO, CTranslate2, Triton or vLLM. As practice, use a relevant one of those tools to compare baseline and quantized inference on latency, throughput and model accuracy; specify a canary check before any production switch.

• Plan deployment and drift response

  The posting calls for real-time inference, Kubernetes, a major cloud provider, monitoring and retraining. As practice, draw a deployment and monitoring plan for one document or face model, including health and model-quality signals, an investigation trigger, and a rollback path. The plan is interview preparation, not a claim about IDfy's internal process.

• Reproduce and simplify a model result

  The role asks engineers to reproduce research and adapt it to production constraints. As practice, select a paper relevant to one listed document or face task, define a reproducible baseline, test one simplification, and present the accuracy-versus-serving-cost result as a model review.

• Select two production-model examples

  The posting asks for at least two models shipped to production. Prepare a one-page summary for each covering your contribution, deployment, monitoring, drift or adversarial issues, and measured outcomes; use only examples you can substantiate.

• Build an end-to-end model walkthrough

  Choose a relevant vision or NLP project and prepare a diagram from data strategy through retraining. Be ready to explain PyTorch training loops, augmentation, loss design, architecture choices, and a training failure you debugged.

• Practice framing a verification problem

  As a preparation exercise, turn the posting’s deepfake-in-KYC example into a defined positive class, objective, error-cost assumptions, operating point, and leakage-free validation plan. Label your assumptions rather than presenting them as IDfy policy.

• Prepare an evaluation worksheet

  Work through a sample confusion matrix, ROC and PR curves, TPR at FPR = 1e-4, calibration, and a prevalence shift. Explain why overall AUC may not describe performance at the chosen operating point.

• Quantify an inference tradeoff

  Prepare a before-and-after comparison for a model optimization, if you have one, covering accuracy, latency, throughput, and serving cost. Be ready to reason about INT8 or BF16, a relevant serving tool, and how a production canary would check accuracy before a GPU-to-CPU move.

• Outline a production operating plan

  Sketch monitoring, drift detection, retraining triggers, and failure handling for a 24x7 model service. Relate the plan to the posting’s real-time scale and be ready to discuss Kubernetes and cloud deployment choices without claiming knowledge of IDfy’s internal setup.

• Prepare research and mentoring examples

  Identify a paper you reproduced or adapted and explain what you kept, simplified, and measured. Prepare a separate example of how you reviewed a model or coached engineers on experimental design and reproducibility.

• Write targeted questions for IDfy

  Ask which document or face model family the hire would own, how operating points and cost-to-serve are decided, what canaries measure, and how model reviews and mentoring work. These are questions, not assumptions about the team’s processes.