What this is
A mock interviewer that can run on a HarmonyOS phone is being developed through 12 tutorials. The author disclosed the current progress: DeepSeek-V3 is used for model calls, the code includes 33 offline test cases, and the system has been divided into five components: persona, workflow orchestration, question bank, multi-turn follow-up questions, and structured scoring.
Across these five layers, the question bank uses RAG, allowing the model to retrieve answers from designated materials before generating a response; workflow is controlled by a state machine; and scoring must produce machine-parsable data. The author also set up a dedicated server to store API keys, interview state, and historical reports, preventing the phone from holding secrets directly or handling complex logic.
Industry view
We notice that this kind of tutorial is closer to what enterprises actually need from AI products than simply demonstrating a chat interface. The model is only one part of the system. Stable workflows, traceable answers, structured outputs, and data security determine whether an application can remain in use. Writing workflow rules into code rather than merely into prompts also shows that vertical agents—AI systems built around specific tasks—are shifting from being able to “talk” to being controllable.
But the risks are equally clear. Whether simulated scores are credible, whether the question bank can cover real job requirements, and whether the model’s follow-up questions introduce misleading signals have not been sufficiently validated. If the system over-scores simply to produce a complete report, users may mistake fluency for improved competence. A human interviewer’s grasp of role context, organizational culture, and live judgment also cannot easily be replaced by a general-purpose model. Moreover, this approach still requires developers to maintain the server themselves, so the actual cost may be higher than the tutorial suggests.
Impact on regular people
For enterprise IT: More applications will adopt a layered architecture in which the frontend handles the experience, while model calls, core workflows, and data remain on the server. API key isolation and traceable results will become basic requirements.
For individual careers: AI mock interviews may lower the cost of practice and provide question-by-question reviews. But they are better suited to identifying gaps in expression and knowledge; they cannot replace job-role matching or human judgment in actual hiring.
For the consumer market: Job-search, assessment, and training applications will move from fixed question banks toward multi-turn conversations. Beyond lower prices, users need to know where the scoring criteria come from, how high the error rate is, and whether the reports genuinely help with career decisions.