KhataPilot

Inspiration

In India, many small shops and traders run on WhatsApp messages written in Hinglish: "Kapoor traders ne 3,000 bhej diye", "10 dabba tel chahiye, Gupta store". The owner keeps the books in their head or in a paper khata, forgets who to chase, and spends the evening reconciling. We wanted a back office that starts from the messages that already exist.

What it does

KhataPilot reads each message with NVIDIA Nemotron on Nebius Token Factory and turns it into a ledger entry (stock, money owed to you, money you owe). Every morning it writes a briefing in English, Hinglish or Hindi: who to chase, what to reorder, what needs a decision. At close of day it freezes balances and writes an Excel-ready day book. Next morning each customer who owes money has a one-click WhatsApp reminder listing their unpaid bills. A person presses send every time. Nothing is sent automatically.

How we built it

The model reads, code decides, you send.

Nemotron-3-Nano-30B-A3B on Token Factory (OpenAI-compatible API) turns a message into JSON: intent, party, items, amount, date.
Code validates every field before the ledger changes. Complaints, unclear messages and payments with no named payer wait in a "needs your decision" queue. Hindi grammar traps ("X ne bhej diye" versus "X ko bhej diye") are guarded in code, and a duplicate message within 10 minutes is held instead of applied twice.
Customer reminders contain money, so they are built by code from the closing balances, with no model involved.
Every number in a model-written briefing is checked against the ledger before it is shown.
Pure Python standard library plus SQLite and a single-file dashboard; it runs on one computer. Only message text and ledger figures for the briefing go to Token Factory.

How we measured it

48 hand-labelled Hinglish messages (28 demo messages plus 20 written after the prompt was frozen), scored by python evalrun.py. Nemotron plus our validation code scored 92% to 98% across four runs (85% to 95% on the 20 fresh messages). Two Hindi grammar guards bring it to 48 of 48 in three runs, but we wrote them after seeing the misses, so we call that in-sample. Switching Nemotron's reasoning off was about 7x faster but fell to 57%, so we keep it on. The dashboard shows per-message latency, tokens and cost per 1,000 messages.

Challenges

Reliability without trusting the model: we found a wrong-direction bug class ("X ka 6,500 pending hai, saturday ko denge") and fixed it with grammar guards rather than a longer prompt.
Briefings can have correct figures but wrong wording. We saw an invented "by tomorrow", a leaked field name, and once a Hindi briefing that reversed who owes whom. The number guard cannot catch those, and we document it openly.
Testing showed that pasting the same payment twice applied it twice, so we added a duplicate guard.

Accomplishments that we're proud of

A trust model a shop owner can understand (the model never writes to the books), a visible trace of how every message was read, and an honest evaluation that reports where it fails.

What we learned

Reasoning matters a lot for Hinglish extraction (57% without it), outputs vary slightly between runs even at temperature 0, and the biggest quality win came from deterministic code around the model, not a bigger prompt.

What's next for KhataPilot

Invoice photos with a vision model, item-to-supplier mapping, the WhatsApp Business API with owner approval, server-side Hindi speech-to-text, fine-tuning on corrected entries, and hosted multi-shop accounts.

Notes for judges

The hosted demo (link below) replays real Nemotron outputs saved from our runs, with no live model calls and no key, so it costs nothing and cannot be abused. Run the repo with your own Token Factory key to read any message live.
Briefings in the demo are real unedited model outputs, chosen from several runs for correctness. Not every run is this clean.
The project was built entirely during the submission period (first commit 2 Oct 2026).
Our feedback on Nebius Token Factory and Nemotron is in FEEDBACK.md in the repo.