ArthNiti
Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

अर्थNiti turns what a first-time borrower says into a plain answer: can this business afford a loan, which government schemes may apply, and what to do next. It is aimed at the lenders and field staff who sit with these borrowers, such as NBFCs, microfinance institutions, self-help-group federations and Common Service Centres. A person speaks or types a few sentences in their own language. Muse Voice Transcribe writes the words down and Muse Spark finds the form answers in them. Every answer is shown with the person's own words beside it, code drops any answer whose words are not in the transcript, and nothing enters the form until the person has seen it. Everything after that is plain Python rules with no model: profit, repayment pressure and the household check, loan sizes with a month-by-month schedule, the schemes that match the work and the place, what each scheme's published terms ask the person to put in and what they add, the registrations to ask about, and district prices where government or NECC data exists. A second, optional use of Muse Spark writes a short plan in the person's language around blanks. It is never shown a figure, a name or a place; the code fills the blanks from the rules and refuses a plan that carries a number of its own, a link or the wrong script. Models do the language; rules do the money. Prior work: the rules engine, the scheme data (30 schemes) and the interface were built before the event, and so was the first version of the intake service (the Meta client and its two routes), written on the morning of 26 Oct, before the kickoff ended at 12:24 PM ET. Built during the event: the voice-and-text page with its recorder and review screen, the spend ledger that keeps the API use inside the free credits, the written plan, the loan options and repayment schedule, the funding card, the registrations card, a green, yellow or red light on the result pages (set by the rules, not by a model), a Simple or Detailed view, egg prices, market rows for 29 more districts, and the scheme list grown from 30 to 99 schemes. The video labels each screen as new or from before the event. Limits, stated on the pages: no scheme's eligibility rules are checked, all figures are estimates and not offers, and the wording in the seven languages other than English is machine-written and has not been reviewed by native speakers. It is a one-person prototype and nobody is using it yet. The submission repository holds the files and changes the event added, not the earlier code: its assemble script rebuilds the full app from one pinned commit of the earlier repository (https://github.com/BU1lDR/ArthNiti), which its README names and explains. A few edits from the 24 minutes before the stated 12:24 PM ET start remain in its patches, and the README names them.