She has cooked for 200 weddings. Give her a certificate.
HUNAR certifies India's 100 million informal workers through voice and live demonstration — entirely in their own language. Hindi, Rajasthani, Bhojpuri. No reading. No typing. No English.
"She has never received a single certificate — not for cooking, not for threading, not for anything. And she has been doing this for fifteen years."
User research · Kota, Rajasthan · 2026
The Problem
100 million workers. Zero proof of skill.
India's informal workforce — cooks, tailors, caregivers, weavers — practise their craft for decades. No platform gives them proof of what they know.
📜
No formal credentials
Informal workers cannot prove skills to employers, even after 20+ years of practice. Without proof, they stay invisible.
🗣️
Language barrier
Every existing certification portal requires reading and typing in English. 78% of India's informal workers are not English-literate. They are invisible by design.
💸
Cost barrier
Training centre certifications cost Rs.500–2,000 — unaffordable for daily-wage workers earning Rs.300/day.
⚡
Gender gap
67% of uncertified informal workers are women, locked out of formal employment and fair wages without recognised credentials.
Multilingual Core
India has 22 official languages. Language is not the barrier anymore.
Every word in HUNAR — every question, instruction, and result — is spoken aloud in the worker's own language. Not translated after the fact. Native from the start.
हिन्दी
Hindi
600M+ speakers · North India
✓ Live
English
English
125M+ speakers · Urban India
✓ Live
राजस्थानी
Rajasthani
80M+ speakers · Rajasthan
✓ Live
भोजपुरी
Bhojpuri
50M+ speakers · UP & Bihar
✓ Live
मराठी
Marathi
83M+ speakers · Maharashtra
Coming next
"I mapped each language to correct BCP-47 codes per Android version — because flutter_tts gives different language codes on Android 11 vs Android 12. Getting this right was non-negotiable."
Real bug fixed during development · See Tech section for details
What Works
Built. Tested. Working.
Here is exactly what HUNAR does right now — not planned, not coming soon. Working today.
Working
Language Selection
Worker selects Hindi, English, Rajasthani, or Bhojpuri. The entire app adapts. flutter_tts reads every word aloud — literacy not required at any point.
Working
14 Profession Categories
Cook, Tailor, Caregiver, Weaver, Construction, Painter, Teacher, Cleaner, and more. Each has custom AI evaluation criteria.
AI Powered
Video Demonstration + AI Scoring
Worker records herself working. Google Gemini evaluates across 7 criteria — Technique, Timing, Hygiene, Ingredient Handling, Presentation. Score: 83/100 demonstrated on real user.
Working
Verified Profile + Marketplace
Worker profile shows verified badge, score, certification level. Employers browse, filter by skill, and contact directly. 11 demo workers pre-loaded for judges.
Home
14 Skills
AI · 83/100
Verified
User Research
Four people. Four skills. One identical answer.
I conducted user research in Kota, Rajasthan with four informal workers. I asked each one the same question: do you have a certificate for your work?
Hostel Warden
Cook + Beauty worker · 15 years
"Nahi. Koi certificate nahi hai." — Her answer when asked if she had a certificate for fifteen years of professional work. She did not even seem surprised.
Changes made after testing with her
Direct product impact from user feedback
She kept looking at question text even though the app was reading it aloud → made text smaller, microphone button much larger. She did not understand "Level 2" → added the word "Proficient".
Mess Cook · Guard · Gym Trainer
Three additional interviews · Same result
Four different skills. Four different people. One identical answer. That pattern — not one person's story — confirmed the problem is systemic, not individual.
AI Evaluation
You cannot fake 20 years of skill in front of a camera.
Google Gemini evaluates the worker's video demonstration across 7 skill-specific criteria. The AI judges technique and knowledge — not camera quality, lighting, or environment. A worker filming on a basic Android in a small kitchen is evaluated on identical criteria to one with a flagship phone.
Technique87%
Timing93%
Ingredient Handling96%
Hygiene70%
Video Quality75%
Presentation75%
Face Detection85%
Technical Work
How I built it.
The key technical decisions — and the problems I had to debug to make them work.
AI Video Evaluation Pipeline
1
Frame extraction — video is sampled at one frame every 2 seconds using the camera package. Experimented with different rates — too many slowed processing, too few missed key moments.
2
Face detection — confirms the same person is present throughout. Confidence threshold set to 0.6 — frames below this are excluded from scoring to handle low-light conditions.
3
Gemini evaluation — frame analysis passed to Google Gemini with a skill-specific prompt. Different criteria weights for Cook vs Tailor vs Caregiver. Returns structured JSON score.
4
Weighted scoring — Video Quality contributes less weight than Technique or Ingredient Handling. A worker with a bad camera but expert technique still scores well.
Inconsistent scores in low light — same worker, same task, score dropped from 87 to 62 in dim lighting. Root cause: face detection confidence below 0.6 throwing off the pipeline. Fix: added minimum confidence threshold, excluded low-confidence frames.
2
Videos not processing on some Android versions — traced to file path issue. External storage paths varied by manufacturer. Fix: switched to app temporary directory, consistent across all Android versions.
3
TTS not reading in correct language — flutter_tts language codes differ by platform. Fix: mapped each of the 5 Indian languages to correct BCP-47 codes per Android version.
4
shared_preferences vs TinyDB — App Inventor V1 used TinyDB with shared namespace. Flutter rebuild uses shared_preferences with consistent key naming across screens for same effect.
GenAI in Development
How I used AI tools to build HUNAR.
I used generative AI tools during development — here are the specific prompts, what they returned, and how I modified the output.
Prompt 1 — Frame extraction logic
"Generate Flutter code to extract frames from a video file at 2-second intervals using the video_player package and save them as image files"
Initial output used deprecated video_player APIs. Modified to use camera package instead for direct recording + frame capture. Had to add error handling for null frames on older Android versions.
Prompt 2 — Gemini scoring prompt
"Write a prompt for Google Gemini to evaluate a cooking video and return a JSON score with criteria: technique, timing, hygiene, ingredient_handling, presentation"
Output was a good starting structure but scores were inconsistent across video qualities. Added explicit instruction: "evaluate technique only, not camera quality or environment" — this made scores consistent regardless of phone quality.
Prompt 3 — App Inventor TinyDB namespace (V1)
"In MIT App Inventor, how do I share TinyDB data between screens without passing values as screen parameters"
Returned the shared namespace solution — set Namespace property to the same value on every screen's TinyDB component. This became the core architecture of V1. Implemented exactly as returned, no modification needed.
Development Journey
From prototype to product.
Two complete versions. Real user feedback between them. Every change has a reason.
TinyDB shared namespace HUNARdb across all screens
Any spoken answer scored a point
5 skill categories
Certificate as text — no QR code
What testing revealed: answering questions about cooking is not the same as cooking. A person can memorise answers. You cannot memorise twenty years of technique. The entire evaluation model was wrong.
→
Version 2 · April 2026
Flutter + Dart + AI
Video demonstration — worker shows her skill on camera
Google Gemini evaluates 7 criteria with percentage scores
Employer voice feedback — no typing needed
14 profession categories
Two-tier certification — Basic and Advanced
Employer marketplace with browse and filter
Score 83/100 demonstrated on real user
The App Inventor version proved the concept. The Flutter version proved the product. Both versions are preserved in the GitHub repository.
Business Model
Three revenue streams. Zero cost to workers.
Rs.40–120
Government Reimbursement
PMKVY already allocates Rs.12,000 crore to certify informal workers. HUNAR applies to be an empanelled agency. Government pays per assessment. Zero selling required.
Rs.500
Employer Placement Fee
Gig platforms pay per HUNAR-certified worker placed. Pre-verified workers reduce complaint rates and onboarding cost. LinkedIn Recruiter model applied to informal labour.
Rs.2k
Institution Subscription
NGOs and training centres pay monthly for a dashboard to issue and track certificates. Proof of impact for donors. Recurring. Low churn.
Year 1 Impact Projections
Based on 5 pilot NGO partners in Rajasthan
10,000
Workers certified
Rs.4L
Govt reimbursement revenue
1,000
Employer placements
Why HUNAR
No one else does all four things at once.
Every existing platform fails at least one of the four things that matter most to an informal worker with no literacy and no money.
Platform
Voice-first
Local language
AI video eval
Free to worker
NSDC / Skill India
✗
✗
✗
✗
WorkIndia / Apna
✗
✗
✗
✓
Traditional training centres
✗
✗
✗
✗
HUNAR
✓
✓
✓
✓
Get Started
Your skill is real. Now it is recognised.
Technovation Girls 2025–2026 · Senior Division · Global Appathon 2026
