MVP Overview · June 2026

Parichay AI

Introduction before the interview

AI screening & proctored assessments for hiring teams. Parichay automates the first gate of hiring with voice + vision intelligence — from AI-generated assessments to live pre-interviews and ranked shortlists.

AI Pre-Interview
Proctored Timed Tests
Ranked Shortlists
6
integrated AI systems
7
step hiring pipeline
3
user personas
The Problem

Hiring teams lose hours before the first real interview

Recruiters spend hours writing assessments and screening unqualified applicants. Phone screens don't scale — they're inconsistent, slow, and hard to record. Remote assessments lack integrity, and hiring teams need signal before investing in full interviews.

  • Hours spent writing assessments and screening unqualified applicants
  • Phone screens don't scale — inconsistent, slow, and hard to record
  • Remote assessments lack integrity: cheating, proxy test-takers, tab switching
  • Hiring teams need signal before investing in full interviews
Our AI-First Solution

Automate the first gate of hiring

Parichay AI automates the first gate of hiring with voice + vision intelligence. OpenAI generates role-specific assessments in seconds, the AI Interviewer conducts live ~4-minute pre-interviews, and on-device ML proctors both phases — no cloud vision API required.

AI Assessment Generation
Live AI Pre-Interview
On-Device Proctoring
Confidence Scoring
Ranked Shortlists
Video Playback
MVP at a Glance

End-to-end hiring, fully integrated

Full hiring pipeline

Job creation → AI pre-interview → proctored assessment → recruiter review — all in one platform.

Six AI systems

OpenAI, AWS Bedrock, Polly, Nova Sonic, on-device ML, and confidence scoring — working together.

Three personas

Recruiter, Candidate (guest apply), and SaaS Owner — each with tailored workflows and permissions.

Platform Architecture

Full-stack, cloud-native hiring platform

Frontend

React + TypeScript
Vite-powered SPA
TanStack Query
API state management
MediaPipe
On-device vision (WASM)
Web Audio API
Mixed session recording

Backend & Cloud

Node.js + Express
REST API layer
Prisma + PostgreSQL
Relational data store
AWS Bedrock
Nova Sonic live interview
OpenAI gpt-4o-mini
Generation & grading
Parichay AI end-to-end hiring pipeline architecture diagram
AI Hiring Pipeline

7 steps from job posting to shortlist

Every candidate moves through a structured, AI-orchestrated journey — from recruiter setup to ranked review.

01
1

Recruiter creates job & assessment

OpenAI drafts questions; recruiter toggles pre-interview on/off for each assessment.

02
2

Polly welcome audio

Neural TTS explains the flow before the candidate joins the live interview session.

03
3

Nova Sonic pre-interview

6 job-aware questions with recording and browser-based proctoring during the live session.

04
4

Confidence gate

Transcript scored on 4 dimensions; unlocks timed test if thresholds are met.

05
5

Proctored timed assessment

Screen/camera recording plus on-device behavior monitor throughout the test.

06
6

Scoring & submission

MCQ auto-grade; OpenAI grades subjective answers; recordings uploaded to S3.

07
7

Recruiter review

Applicants list, video playback, and rankings ordered by assessment score.

AI Systems Deep Dive

Cloud AI + on-device ML

Six integrated AI systems power every stage of the hiring pipeline — from question generation to proctoring.

OpenAI — Assessment Generator

Generative AI · Recruiter-facing

Generates role-specific assessments in seconds based on job title, role, and experience level.

  • Model: gpt-4o-mini (configurable via OPENAI_MODEL)
  • Output: ~10 questions — MCQ, coding, and subjective
  • Fallback: deterministic template bank when no API key is configured
  • Recruiters toggle "requires pre-interview" on the same assessment
Proctoring

Integrity at every stage

Pre-Interview Proctoring

Multi-face detection (2 strikes → reject), camera obstruction, dim lighting, head movement & gaze away. Nova reads alerts aloud; mic paused during warning playback. WebM records mic + AI audio mixed → S3.

Timed Assessment Proctoring

On-device ML monitors lighting, head turns, gaze, multiple faces, and movement. Tab switch, window blur, or fullscreen exit → 3 total violations auto-submit. Screen share + mic OR camera + mic WebM at VP9/Opus, capped 480p/720p @ 15fps.

Recruiter Workspace

Everything recruiters need in one dashboard

01

AI assessment generation

Generate assessments from job role and seniority in seconds.

02

Applicants management

Jobs table with per-job applicants modal — confidence, pass/fail, recordings.

03

Video playback

Pre-interview & assessment video playback from S3, company-scoped.

04

Rankings & scoring

Completed assessments ordered by score with attempt detail and AI feedback.

05

Company settings

Cooldown management and pre-interview toggle per assessment.

06

Violation metadata

Full attempt detail: answers, AI feedback, transcript, and violation logs.

Shipped MVP

Production-ready capabilities

  • Guest apply flow (/apply) — no account required
  • Optional Nova pre-interview per assessment (toggle on create or inline)
  • 90-day company cooldown after rejection or completion
  • Mixed session recording (candidate mic + Nova audio in one WebM)
  • Live camera preview + mic indicator during timed test
  • Submission safeguard: fullscreen loader during answer + video upload
  • AI overview page for candidates explaining the pipeline
Policy Thresholds

Clear limits for fairness & integrity

Pre-interview limits

3 audio interruptions, 2 multi-face, 2 camera obstruction, 5 behavior strikes. ~4 min / 6 questions, 8 min max WebSocket connection.

Confidence gate

Overall ≥ 60, Communication & Job fit ≥ 70% of dimension. Fail → ineligible with 90-day company cooldown.

Assessment rules

3 integrity violations → auto-submit. 75% pass threshold. 10 marks per question. 90-day re-application cooldown per company.

AI Engineering

Technical highlights

Nova Sonic bridge

Bidirectional stream bridge with proctor hold & playback interrupt. PROCTOR_ANNOUNCE pattern avoids Bedrock duplicate SYSTEM content error.

Shared behavior monitor

One on-device behavior monitor for pre-interview and assessment phases with baseline calibration (~14 samples) and hysteresis.

Web Audio mixer

Routes Nova playback to speakers AND recording track. Mixed WebM: mic + AI audio in a single upload.

Graceful fallbacks

OpenAI + heuristic fallbacks when API keys are not configured. Deterministic template banks for assessments.

Demo Walkthrough

See the full pipeline in action

1

Recruiter: Create job → Generate assessment (OpenAI) → Enable pre-interview → Share link

2

Candidate: Enter details → Polly welcome → Nova Sonic interview → Confidence result

3

Candidate: Timed proctored test → Submit (auto-grade + AI subjective feedback)

4

Recruiter: View applicants → Play recordings → Rankings

5

Direct-assessment path skips Polly + Nova when pre-interview is off

Introduction before the interview

AI screening & proctored assessments — ranked shortlists, confidence scores, and video playback.

Parichay AI by Vanshiv · MVP June 2026