Building Subhux HireUp: AI in Recruiting

At Subhx Infotech, we engineered an automatic applicant evaluation platform. The core goal was to conduct technical screening rounds asynchronously while maintaining proctoring metrics.

Architecture Pipeline

The system is split into three core phases:

1. Client Capture: WebRTC media recording captures the user's screen and webcam.

2. Asynchronous Upload: Media streams are uploaded directly to secure storage endpoints.

3. LLM Structured Audit: Transcripts and activity logs are parsed through Google Gemini models.

[Candidate UI] --(WebRTC)--> [S3 Upload] --> [Redis Queue]
                                                    |
                                             [Worker Nodes]
                                                    |
                                            [Gemini Assessment]

Optimizing LLM Evaluations

To get structured, non-biased assessments, we avoid open-ended prompts. Instead, we use Gemini's structured response schema feature:

const evaluationSchema = {
  type: "object",
  properties: {
    technicalScore: { type: "number" },
    communicationScore: { type: "number" },
    reasoningExplanation: { type: "string" },
    recommendedQuestions: { type: "array", items: { type: "string" } }
  },
  required: ["technicalScore", "communicationScore", "reasoningExplanation"]
};

Preventing Proctoring Cheats

To prevent browser-switching and multi-screen cheating, we listen to focus events:

window.addEventListener("blur", () => {
  logSecurityIncident("Candidate switched windows / tab");
});

This combination of client-side tracking and AI-driven analysis successfully cut overall technical filtering workloads by 75%.