How to write it
Own the serving path, and say so
The clearest way to establish that you are an engineer rather than an analyst is to describe the production system: throughput, latency, availability, and what depends on it. "8k predictions/second, p99 40ms, driving 11% of site revenue" is a sentence a notebook-only candidate cannot write.
This is also the answer to the most common screening question about ML resumes — did any of this ever run? Lead with the fact that it did.
Pipelines beat models
A single well-trained model is a one-off. A pipeline that retrains, evaluates, and promotes models without a human is infrastructure, and it's what companies are short of. Describe the machinery: feature computation, training orchestration, evaluation gates, promotion, rollback.
The metric to attach is cadence. "Model refresh from quarterly and manual to daily" tells a hiring manager exactly what capability you added, and implies everything about the engineering underneath it.
- Retraining frequency, before and after automation
- Experiment turnaround time
- Predictions per second and serving latency
- Inference or training cost, and how you reduced it
- Drift or degradation caught before customers noticed
Evaluation is the credibility centre of an ML resume
Anyone can report a metric on a test set. The engineer worth hiring is the one who has made offline evaluation actually predict online behaviour, built golden sets, gated deploys on regression tests, and caught silent failures in production.
"Rebuilt offline evaluation to match online results within 2%, ending a long-standing trust problem with launches" is a strong bullet precisely because it's about the discipline of measurement rather than about a model architecture. Hiring managers who have been burned recognise it immediately.
Cost is now part of the job
Inference cost has become a first-order constraint, especially anywhere large models are involved. Distillation, quantisation, batching, caching, moving off GPUs where CPU will do — these are ordinary ML engineering decisions now, and quantifying one puts you in a small minority of applicants.
State it as a percentage or a monthly figure and name the mechanism. "Cut inference cost 58% by distilling the ranking model and moving serving to ONNX on CPU" is engineering, finance and judgement in one line.
If you work with LLMs, be concrete about what you built
"Experience with LLMs" and "prompt engineering" have been diluted to the point of meaning nothing. What still signals competence is the surrounding engineering: evaluation harnesses, retrieval pipelines with measured relevance, guardrails, latency and cost budgets, caching strategies, and regression gates for prompt changes.
Name the system and its constraint. An LLM evaluation harness with golden sets and CI gates is a real artefact; a list of model names you've called an API for is not.