A one-screen tool for the permission clerk: reads a question in English, Hindi or Hinglish, calculates setbacks, FAR and fees in code, and cites the exact rule.
The problem
A building-permission clerk at the Bhopal Municipal Corporation answers the same questions all day: "My plot is 200 sq m on a 9 m road. How much gap do I leave at the front, how many floors can I build, and what fee do I pay?" The answers sit in the MP Bhumi Vikas Rules 2012: 108 pages, 105 rules, with setback, FAR and fee numbers spread across tables that are easy to misread. The rules keep changing (consolidated to March 2025, amended again on 28 August 2026), and questions come in English and Hindi. The clerk needs a correct answer within a minute, with the rule number to show the citizen. Today a misread table means a citizen pays fees and submits plans, and only later learns the plan breaks a rule.
How it works
Kshiti is a one-screen tool for the clerk. The clerk types the citizen's question in English, Hindi or Hinglish, and Kshiti:
- Reads the question in under a second: language, topic and the citizen's numbers (road width, plot area, height). Any number not actually written in the question is discarded.
- Calculates figures in code, not with AI: front, rear and side open space, FAR and maximum built-up area, height limit and fees. They come from tables checked against the official PDF, with the exact table row shown.
- Writes a short answer in the question's language using Claude Opus 5.5, from the 106 rules only, citing the exact paragraph.
- Checks every quote word for word against the official text. Quotes that don't match are removed and made-up rule numbers are flagged. If the calculator has a verified figure, the figure always wins.
- Says "not found" clearly when the rules are silent, and asks for missing facts instead of assuming them.
- Shows the version of the rules used (consolidated to 06-03-2025, with a SHA-256 fingerprint) and flags the 28-08-2026 amendment to Rule 2(5)(c).
- Prints a bilingual PDF answer slip for the citizen: rule, official text, receipt ID and QR code.
Measured on the 17 official sample questions (Claude Opus 5.5): correct rule 15/15, "not found" 2/2 (0 wrong), table figures 5/5, answer language 17/17, quotes verified word for word 107/107, invented rules 0, amendment flagged 1/1. Median time 12.8 s, with the figures appearing instantly.
Known limits (also in the README): the calculator covers residential buildings only; quotes come from the English PDF because the Hindi PDF can't be machine-read; only the 28-08-2026 amendment is known. Kshiti points to the rule; the clerk decides.
How Claude was used
Claude Opus 5.5 is the answer engine. The whole rulebook is split into 867 quotable blocks, and each of the 106 rules is sent as a citable document with built-in citations, so every quote maps back to a rule and PDF page. The rulebook is prompt-cached (about 113k tokens read from cache per question). Tools handle "not found" and "ask for missing inputs", and answers stream in. The calculator's figures are given to Claude so its text matches the code, and every citation is then verified word for word.
The whole project was built with Claude Code during the event, from 12:58 PM: planning, data checks against the official PDF, the calculator and tests, the interface, the PDF template, the evaluation on the 17 sample questions, bug fixes found by that evaluation, and the Vercel deployment. The commit history in the repo shows each phase.
Team
THETA1728