Review the evidence signals before interviewing. Then use the anchored descriptions—not instinct alone—to choose the score that best matches each answer.
01
Evaluation factor
Technical proficiency
35% weight
Check depth on distributed training stacks: PyTorch DDP or FSDP, NCCL collectives, CUDA memory profiling, Kubernetes plus Kubeflow or Ray, and Triton or vLLM serving.
Evidence to listen for
Command of the languages, frameworks, and data tools the role actually uses
Understands correctness, performance, and failure modes, not just syntax
Has opinions on testing and can justify them
Reads and reasons about code they did not write
Five-point scoring guide
1
Poor
Cannot work independently; fundamentals are missing.
2
Needs Improvement
Weak fundamentals; output needs heavy review.
3
Satisfactory
Competent for the role; needs guidance on complex or unfamiliar work.
4
Very Good
Strong practitioner; handles hard problems with little guidance.
5
Excellent
Names specific frameworks and versions, explains NCCL ring topology or FSDP sharding trade-offs, and quantifies GPU utilisation gains achieved.
02
Evaluation factor
Systems and trade-offs
25% weight
Probe how they sized clusters and pipelines: spot versus reserved GPU capacity, checkpointing strategy, feature store design, and cost per training run versus latency targets.
Evidence to listen for
Reasons about scale, latency, cost, and failure before writing code
Names the trade-off they chose and what they gave up
Understands the data lifecycle end to end
Anticipates what breaks at ten times the volume
Five-point scoring guide
1
Poor
No thinking beyond the immediate task; no awareness of scale or failure.
2
Needs Improvement
Limited architectural awareness; struggles with design decisions.
3
Satisfactory
Works within an existing design; makes sound local decisions.
4
Very Good
Designs for scale and maintainability; articulates trade-offs clearly.
5
Excellent
Walks through a real capacity decision, states the cost and reliability trade-off, and explains what they deliberately chose not to build.
03
Evaluation factor
Evidence and rigour
25% weight
Test measurement habits: MFU and throughput benchmarking, straggler detection, canary model rollouts, drift monitoring in Prometheus or Grafana, and reproducibility of training runs.
Evidence to listen for
Validates results rather than trusting output
Knows how their work is measured and what a bad result looks like
Can describe a time their own analysis or model was wrong and how they caught it
Careful about data quality, leakage, and silent failure
Five-point scoring guide
1
Poor
Ships unvalidated work; no notion of how correctness is checked.
2
Needs Improvement
Validates superficially; misses obvious quality or leakage issues.
3
Satisfactory
Reasonable checks in place; rigour drops under time pressure.
4
Very Good
Validates thoroughly; can name a real error they caught in their own work.
5
Excellent
Cites before and after numbers (tokens per second, p99 inference latency, failed job rate) and describes how runs were made reproducible.
04
Evaluation factor
Collaboration and communication
15% weight
Assess how they support research users: on-call for training job failures, internal platform docs, runbooks, and negotiating GPU quota between competing model teams.
Evidence to listen for
Explains technical work to non-technical stakeholders
Gives and takes code or peer review constructively
Documents enough that the work survives their absence
Aligns with team process rather than working around it
Five-point scoring guide
1
Poor
Cannot work in a team; resistant to feedback.
2
Needs Improvement
Communication issues create rework; lone-wolf tendencies.
3
Satisfactory
Adequate team member; documentation and review participation are light.
4
Very Good
Communicates well; reliable reviewer and collaborator.
5
Excellent
Describes concrete platform adoption, gives examples of unblocking researchers quickly, and shows fair, documented handling of contested GPU allocation.
Put this rubric to work
Score every candidate against the same standard
Add these weighted factors to Hirevire and let AI evaluate recorded answers against your rubric.
