KaushalSetu (Autonomous Vocational Taskmaster)

KaushalSetu is an autonomous vocational AI agent that auto-synthesizes curricula, evaluates multi-modal exams, mints tamper-proof SHA-256 ledger seals, and matches students with live regional jobs.

Comment

Inspiration

I run a small computer coaching center, and I live this reality every single day.

After every batch wraps up, a grueling cycle begins: writing test papers by hand, designing updated curricula, grading answer sheets, and calculating transcripts one by one. But the hardest part hits after students graduate. Small institutes do not have dedicated placement departments or large HR teams. I found myself spending endless late nights sifting through job portals, trying to match individual resumes to genuine regional vacancies, conducting mock interviews, and cold-emailing companies for each candidate.

When I first learned about this Google competition, I spent days brainstorming theoretical ideas. Then it clicked: why look for an abstract problem when the biggest, most exhausting bottleneck was sitting right in front of me at my own center? For years, I had normalized losing weeks of productive time to repetitive administrative chores without realizing just how much momentum we were losing.

I realized: why not build an autonomous AI agent to solve our own daily struggle? An engine capable of turning raw notes into standardized curricula, grading exams objectively, minting tamper-proof credentials, finding verified live jobs, and coaching students through interactive voice interviews. Huge thanks to Google for organizing this competition—it gave us the exact catalyst to turn our personal ground-level challenge into KaushalSetu.

What it does

KaushalSetu operates as an Autonomous Multi-Tenant AI Taskmaster that bridges the operational divide between grassroots vocational institutes, students, and employers:

  • Instant Curriculum Auto-Synthesis: An instructor enters raw, rough keywords (such as ev battery mechatronics), and the agent automatically synthesizes an industry-aligned 4-week modular syllabus, adaptive multimodal MCQs, and practical diagnostic capstone briefs.
  • Dual-Pillar Multimodal Grading: The engine evaluates candidate submissions across theoretical MCQs ($50$ marks) and practical telemetry/code assignments ($50$ marks), computing a transparent aggregate score: $$\text{Aggregate Score } (S) = M_{\text{theory}} + C_{\text{practical}} \quad \text{where } S \in [0, 100]$$
  • Cryptographic Tamper-Proof Seal: Upon exam completion, the platform mints an immutable mathematical SHA-256 ledger seal: $$\text{Seal} = \text{0x} + \text{SHA-256}(S_{\text{id}} \parallel S \parallel T_{\text{timestamp}})$$ Any employer can paste this seal or the Student ID into our Public Verifier to validate authenticity in seconds, eliminating credential fraud.
  • Dynamic Glassmorphic Portfolio: Automatically compiles the student's profile photo, verified skills, and GitHub repositories into a modern, responsive web portfolio landing page.
  • Live Job Intelligence & 1-Click Outbox: Aggregates verified openings from the National Career Service (NCS) and live industry feeds, highlights roles with a "Top 2 Highest Selection Probability" badge, and logs 1-click candidate applications directly into the institute's mentorship outbox.
  • Turn-by-Turn Conversational Voice Interview Studio: Rather than a static quiz, the AI conducts an interactive, multi-turn technical interview where students can speak their responses using the browser Web Speech API while receiving instant diagnostic marks and feedback per turn.

How we built it

  • Hybrid Architecture: Built using an asynchronous FastAPI backend paired with a mobile-first, reactive Streamlit user interface.
  • Zero-IPC In-Process Execution: Critical grading logic, state transitions, and ledger minting execute in-memory with sub-$10\text{ms}$ latency, eliminating localhost network bottlenecks.
  • AI & Reasoning Core: Powered by Google Gemini 2.5 and Gemma Token Screener models for curriculum synthesis, multimodal rubric grading, and dynamic interview probing, supported by deterministic heuristic fallbacks for uninterrupted reliability.
  • Storage & Schema Integrity: Implemented SQLite with Write-Ahead Logging (WAL) and dynamic self-healing column migrations (PRAGMA table_info) to prevent runtime database locks and column drift.

Challenges we ran into

  • Multi-Device Responsiveness: Making complex dashboards, transcript previews, and voice interview stations render cleanly on mobile phones required custom CSS grid and flexbox breakpoints to ensure multi-column layouts stack naturally on smaller screens.
  • Session State Management: In a multi-tenant center roster, switching between multiple candidates was causing Streamlit session cache bleeding. We solved this by engineering an explicit state router with dedicated cache-purging routines upon every profile launch.
  • Zero-Failure Live Demo Architecture: To ensure the system never breaks due to external API rate limits or network drops during live presentations, we engineered dual-track execution pipelines where Gemini calls fail over gracefully to local heuristic rule engines.

Accomplishments that we're proud of

  • 15 Days of Work in 30 Seconds: Administrative workflows that previously consumed nearly two weeks per batch (curriculum planning, exam paper setting, marksheet calculation, and job matching) now run autonomously in under 30 seconds.
  • Enterprise Trust for Grassroots Centers: We brought corporate-grade cryptographic verification to local vocational certificates.
  • Functional Voice Agent: Enabling students to dictate their technical methodology directly through their device microphone and receive instant AI coaching was our biggest engineering milestone.

What we learned

  • Practical AI value does not come from building simple conversational wrappers, but from eliminating deep manual friction in daily operations.
  • Production systems require deterministic architectural guardrails, solid database schema migrations, and low-latency in-process execution far more than complex prompting.

What's next for KaushalSetu (Autonomous Vocational Taskmaster)

  • Direct automated WhatsApp and SMS notification bots for candidate interview reminders.
  • Offline edge model inference for rural training centers operating on intermittent internet.
  • Direct API integration with state skill registries and India's national DigiLocker ecosystem for transferable vocational credits.

Built With

  • css3
  • fastapi
  • gemini-api
  • gemma
  • google-cloud-run
  • google-gemini
  • html5
  • pydantic
  • python
  • sha-256
  • sqlite
  • streamlit
  • web-speech-api
Share this project:

Updates

Submission history