Menu
Let’s talk

Open to consulting & senior roles

← All work

AI · GovTech · 2026

Sevak AI: ask any government-scheme question, in your language

A news and government-scheme platform in Marathi, Hindi and English, with Sevak AI — an assistant that answers “which scheme can I apply for?” from a crawled, vector-indexed knowledge base.

Visit mukhyamantri.com/chat
Client
Mukhyamantri.com
Role
Solo — architecture to deploy
Stack
Django · FastAPI · Celery · Next.js · Postgres + pgvector · Redis · Vault · bge-m3
CitizenNext.js (3 lang)Django+FastAPIRule+LLM routerpgvectorLLMCrawlerRedis cache

Simplified architecture sketch

  • 3languages, with hreflang SEO
  • RAGrule + LLM router, streamed answers
  • Solodesign, build, infra, deploy

Context

Information about state and central government schemes is scattered, mostly in English, and hard to search. Most of the people it is meant for search in Marathi or Hindi.

The challenge

  1. 01Answer eligibility questions accurately, in the user’s language, without hallucinating schemes.
  2. 02Keep answers fast and cheap enough to run as a public service.
  3. 03Rank in search across three languages.

Approach

Crawl, chunk, embed

Celery crawlers pull scheme data; bge-m3 multilingual embeddings in Postgres + pgvector power retrieval across Marathi, Hindi and English.

Route before you generate

A rule + LLM router sends easy questions to deterministic answers and only hard ones to the model; answers stream with per-stage timings.

Tolerant search

Typo-tolerant matching for transliterated queries, with a Redis cache in front of retrieval.

SEO in three languages

Next.js with hreflang alternates and PageSpeed tuning; secrets in HashiCorp Vault; Cloudflare tunnel in front.

Result

  • ✦Live at mukhyamantri.com with Sevak AI answering scheme questions in three languages.
  • ✦Benchmarked retrieval pipeline with per-stage latency, versioned through several iterations.

My own product — designed, built and run solo.