Nemotron reads
It turns Hinglish like "Kapoor traders ne 3,000 bhej diye" into structured data: who, what, how much, by when. That is all it does.
Your shop already runs on WhatsApp.
KhataPilot reads the messages, keeps the books, and gets tomorrow's reminders ready. You press send.
Tap a message. Watch it become a ledger entry.
These are real answers from Nemotron on Nebius, saved from our test runs. Nothing is sent anywhere.
Nemotron read
Code checked
The model reads. Code decides. You send.
Code decides
Every field is validated before the ledger changes. Unclear messages, complaints and unknown payers wait in a queue for you.
You press send
Reminders list each unpaid bill and the total. Code writes the money, never the model. WhatsApp opens with the text filled in.
Measured, and honest about it.
92 to 98%
of 48 hand-labelled Hinglish messages read exactly right by Nemotron plus our validation code, across four runs.
57%
is what it scores with Nemotron's reasoning switched off. About 7x faster, far less reliable. We keep it on.
0
messages sent automatically. A person presses send on every reminder and every supplier order.
Where it misses. Almost always a wrong direction on messages like "X ka 6,500 pending hai, saturday ko denge". For a customer already in the ledger, code corrects it from their record.
What we won't claim. Two grammar guards bring the score to 48 of 48, but we wrote them after seeing those misses, so that number is partly in-sample. The model-alone range above is the honest one. 20 of the 48 messages were written after the prompt was frozen; the model scored 85% to 95% on those.
Hinglish is the hardest case.
Shopkeepers write the way they talk, mixing Hindi and English in Roman script. The same briefing comes out in whichever language the owner prefers, and every figure is checked against the ledger before it is shown.
Written by Nemotron from a demo ledger. Every number is verified against the books.
Questions we expect.
What stops the model from corrupting the books?
The model never writes to the database. It returns JSON, and code validates it field by field. Anything unclear goes to a queue for the owner. Reminder text is built by code, and every number in a model-written briefing is checked against the ledger.
What if it is wrong but looks valid?
That is the real residual risk, so we show our miss rate instead of hiding it. The owner can open the raw model output for every message. Known Hindi traps are guarded in code, a known customer's record overrides a wrong direction, and the same message twice in ten minutes is held instead of applied twice.
Why not Tally, Vyapar or a spreadsheet?
Those need the owner to type every entry. KhataPilot starts from the messages that already exist, then handles the follow-up: who to chase, what to reorder.
Is WhatsApp integrated?
On purpose, not through the Business API. One-click links keep a person on every send, which is the safety design, and avoid bulk-messaging risk.
Where does my data live?
On your own computer, in a SQLite file. Only message text and ledger figures for the briefing go to Nebius Token Factory. There is no cloud database and no login, so it is built for one shop on one machine today.
Why is it not instant?
Reasoning is on, which is what makes Hinglish reliable. A message typically takes a few seconds. Messages are independent, so they can run in parallel.