VidyutMitra

01The problem

Help MESCOM households uncover the money they quietly lose to bill traps (Fixed Charge over-provisioning, Gruha Jyothi cliffs, and missed solar subsidies) through a zero-typing, zero-literacy WhatsApp advisor in their own language.

22.6 lakh MESCOM households pay an electricity bill every month with no idea they are losing money to traps the bill never explains. The KERC tariff order runs 47 pages, bills mix Kannada and English, and many consumers cannot read either fluently.

02What I built

A WhatsApp-first pipeline: Gemini 2.5 Flash extracts every bill field, three deterministic Python modules run the math in parallel, and Groq Llama 3.3 70B adds only two personalized lines, validated verbatim against the analysis. Sarvam synthesizes a Kannada voice note and Twilio dispatches text, voice, and infographic in ~12 seconds.

• One photo of a bill returns a full analysis in 10–15s: a text report, a Kannada voice note, and a printable infographic, no typing, no app install
• Triple deterministic analysis per bill: Tariff Engine (Fixed Charge Trap), Subsidy Navigator (PM Surya Ghar & Gruha Jyothi), and a Solar ROI calculator
• Voice-first Kannada output via Sarvam's native TTS, making the analysis accessible regardless of literacy
• Hallucination-guarded AI: every number in an AI-framed reply is regex-validated verbatim against a deterministic fact-pack before it sends
• DPDPA-compliant by design: bill image bytes are never written to disk and the schema has no image column
• Two-turn follow-up drilldowns (SOLAR / FIXED / SUBSIDY / CLIFF) served from a 30-minute cached analysis

03The hard part

Vision, three analyses, an LLM and a Kannada voice note, all back as one WhatsApp reply in under 15 seconds.

04Results

Saved / Example Household

₹1,740/yr

At MESCOM Scale, Projected

₹39 Cr/yr

Cost / Bill

~₹1.7

05Stack

• Python
• FastAPI
• Gemini 2.5 Flash
• Groq Llama 3.3 70B
• Sarvam AI (Kannada TTS)
• Supabase
• Twilio WhatsApp
• Next.js