A real mock interview report, walked through in English

Question

Please introduce yourself, and describe your most relevant hands-on experience for this role's focus on an enterprise knowledge base and disaster-recovery operations.

What the candidate said — Original (Chinese, verbatim)

「我看这个岗位是JD里企业知识库和运维栽培的场景跟我的经验对得上。我既有RAG落地,也有私有化部署。我在两家公司都是从零把AI能力做到上线,来你们这边做运维栽培方向的大模型落地,节奏是我熟悉的。」

English translation

“As I see it, the enterprise knowledge base and “ops cultivation” scenarios in the JD match my experience. I've shipped RAG, and I've done on-premises deployments. At two companies I took AI capabilities from zero to launch, so delivering large models for “ops cultivation” here is a pace I know well.”

The problem

The candidate shows a serious lack of understanding of the core business scenario, misreading “ops disaster recovery” as “ops cultivation”, and stays at slogan-level claims without giving any concrete approach to the core steps — knowledge-base chunking, retrieval, evaluation or on-premises deployment.

A better answer

First get the business right. Then, for data-centre disaster recovery and operations inspection, state the concrete engineering work and metrics from your on-premises RAG pipeline — for example document splitting and cleaning, hybrid vector and full-text retrieval, and long-context handling on self-hosted open-source models.

Question

Please go into a representative result you delivered in model fine-tuning or quality optimisation, and the metrics used to evaluate it.

What the candidate said — Original (Chinese, verbatim)

「漏洞检测项目,我们标注了六万多条数据,用Lora微调大模型,把静态分析误报率从65%降到25%。另外结合AST和CFG图结构,做代码审计,多任务指令微调加思维链,内部测试准确率到98%。」

English translation

“On the vulnerability-detection project we labelled more than 60,000 samples and fine-tuned a large model with LoRA, cutting the static-analysis false-positive rate from 65% to 25%. We also combined AST and CFG graph structures for code auditing, with multi-task instruction tuning plus chain-of-thought, and reached 98% accuracy in internal testing.”

The problem

The figures given (98% test accuracy for code auditing) are far from what is realistic in industry, with no description of how the evaluation set was built, no baseline comparison and no analysis of real failure cases — clear signs of resume padding and buzzword stacking.

A better answer

State plainly how large the evaluation set was, how the benchmark was set up and which metrics you used (precision, recall, F1), and walk through what actually went wrong on long-tail vulnerabilities or hallucinations and how you guarded against it — instead of quoting an unrealistically high accuracy figure.