JobSphere brings multilingual AI guidance to government employment platforms
Government employment sites do a lot, but users still get stuck: complex menus, limited language options, and little personal guidance. JobSphere addresses that gap for the Punjab government's PGRKAM platform with an AI career copilot that speaks English, Hindi, and Punjabi - with voice support.
Created by Srihari R, Adarsha B V, and Mohammed Usman Hussain with colleagues at Presidency University, the system focuses on accuracy, speed, and cost. It runs on consumer-grade GPUs using 4-bit quantization, which cuts costs by 89% versus cloud-heavy deployments while keeping performance tight.
Built for PGRKAM, built for access
- Multilingual assistance: English, Hindi, Punjabi - including voice
- Responses grounded in verified documents via RAG for higher trust
- Median response time: 1.8 seconds
- 94% factual accuracy in evaluations
- 50% improvement on the System Usability Scale compared to the existing PGRKAM experience
How it works (in plain terms)
JobSphere uses the Llama 3.2 3B model with 4-bit quantization to reduce memory needs and run smoothly on readily available hardware. Retrieval-Augmented Generation keeps answers aligned with official sources, which helps reduce misinformation and support consistent service.
- Mock tests are auto-generated from past papers using NER, POS tagging, and syntactic parsing, hitting 85% accuracy for relevant extraction.
- Resume parsing turns unstructured CVs into structured profiles with 92% accuracy, using contextual embeddings and machine learning classifiers.
- Job recommendations use a two-stage setup: embedding translation with cosine similarity and a gradient-boosted tree for ranking.
- Listings stay current through managed scraping with Selenium and BeautifulSoup.
- Stack choices: FastAPI (async APIs), Pydantic validation, JWT auth, React 18, Vite, TailwindCSS - plus data structures like B-trees and hash tables for performance.
Why this matters for government teams
- Reach more citizens: language and voice support reduce drop-offs and help first-time users.
- Lower operational spend: consumer GPUs and 4-bit quantization keep budgets under control.
- Trust by design: answers tied to verified documents improve confidence and consistency.
- Service quality: sub-2 second replies keep users engaged and reduce support load.
- Practical deployment: feasible in on-prem or state data center setups.
What you can do next
- Run a pilot on a single portal or region using a small GPU node; track SUS and response times.
- Curate a clean document base (eligibility rules, FAQs, circulars, step-by-step workflows) for grounding.
- Set guardrails early: JWT-based access, audit logs, fallbacks for edge cases, and clear escalation paths.
- Plan upgrades: expand the mock test library and refine resume parsing for local formats and languages.
Results at a glance
- 89% cost reduction vs cloud-heavy deployments
- 94% factual accuracy
- 1.8s median response time
- 50% usability improvement (SUS) vs current PGRKAM platform
For technical details and evaluation data, see the preprint: JobSphere: An AI-Powered Multilingual Career Copilot for Government Employment Platforms.
If you're organizing internal upskilling around RAG systems and practical deployment, explore focused learning paths here: AI courses by job.
Your membership also unlocks:
