11 September 2026 · २०८३ भदौ २६ · 7 minutes
How Likhat AI works, from a brief to a Nepali draft
The first question a partner asks us is short. Is this just ChatGPT?
The answer is no, and the useful version takes a few minutes. This piece walks the whole run, from the two sentences a lawyer types to the file that comes back.
Likhat AI, made by Hardflip Technologies, is a drafting tool for Nepali advocates that turns a short brief into a first draft in Nepali with statute citations resolved against a real corpus. A lawyer reviews, a lawyer signs.
Two words first.
A model is a program that writes text by predicting the next word. It can write very well. It can also be confidently wrong, because a wrong sentence and a right one look the same to it.
Code is a set of fixed rules. Given the same input it does the same thing every time. It does not write well. It does not need to.
Likhat is built out of both, and it is one system, not two products stitched together. The stages below are all Likhat. A model is used inside some of them.
One: it reads the brief
You give it a short brief. Two or three sentences is normal. You can attach documents instead, and it will read them.
Likhat pulls out the parties, the dates, the property, the amounts. It shows you what it found and, more usefully, what it did not find. Missing facts are listed rather than guessed.
Two: it works out what document this is
This is the step people do not expect, and it carries the rest.
A gift deed is not a sale deed. A bail petition is not a writ. Each one has its own required contents, its own forum, its own governing sections. So Likhat decides which of these you are asking for, or uses the type you chose.
Get this wrong and everything after it is wrong in a tidy, convincing way. So Likhat refuses the job when it cannot work out the document and the lawyer has not chosen one. It does not draft something plausible instead.
Three: it fetches the law itself
Likhat holds a library of Nepali law as files: acts, rules, codes, the राजपत्र. It does not search the open internet, and it does not ask the model what the law says.
For the document type it settled on, it opens the governing acts, slices out the relevant sections, and assembles them. The drafting stage receives that text.
This is the sentence most of our design rests on. The law that reaches the draft came out of a file, not out of a model's memory. What that guarantees, and what it does not, is set out in why a draft cannot cite a section that does not exist.
Four: the Nepali gets written
Now the writing. A Claude model from Anthropic writes the Nepali, working from the facts Likhat extracted and the sections Likhat supplied.
It is a drafting stage, not a chat. It runs apart from our own settings and instruction files, so chat-style instructions cannot leak into a drafting job.
Five: Likhat checks what came back
The draft does not go to the lawyer as written. Likhat's own checks run over it, each doing one job, in order, in code.
The citation check. Every place the draft names a law, the check asks whether that law matched a real file. If it did not, the citation comes out of the text and goes into a list for the lawyer.
The section check. Where the law did match a file, this check opens the file and looks for the section number. If the section is not in it, the number comes out.
The date check. Nepali deeds carry dates in two calendars. This check compares each pair. The Bikram Sambat date is the one that counts, and a wrong AD date beside it is worked out again. Sometimes a date the matter already stated gets retyped. The matter's own date is then put back. The change is flagged in Nepali, so the lawyer sees it.
The personal-details check. Names, citizenship numbers and similar details are replaced with blanks that the lawyer fills in before the draft is shown.
The warning line. Every draft gets a do-not-file line in its own body: "AI-generated draft — not lawyer-verified. Review every citation, fact, and clause before relying on it. Do not file without a licensed advocate's review." It is in the web view, in print, and in the raw file.
There are more checks than these. They work the same way: read the text, compare it against something fixed, change or flag.
Why the checks are code and not instructions
A model can be talked into things. That is not a flaw to be patched. It is how the thing works. It follows instructions, and an instruction is a request among other requests.
So a rule that lives in an instruction is a rule that can lose an argument. A rule that lives in code cannot lose one. It does not read the draft's tone. It has no view about whether this clause is important. It compares a value against a file and acts.
That is why every rule we actually rely on has been moved into code.
What actually stops a job
Partners expect a long list here. There are two things.
A brief with no parties, no facts and no text stops before any draft exists. A type of paper Likhat could not work out, and the lawyer did not choose, is refused.
Everything else flags or amends. Missing facts do not stop a job, because a first draft the lawyer completes is the product. The job of the checks is to keep unsupported text out of the draft, not to withhold the draft.
That is a deliberate choice and it has a cost. A thin brief still produces a draft, and that draft has more blanks in it. We would rather hand a lawyer a draft with 9 blanks than refuse the matter.
What you get back
A draft in Nepali. The law sections in numbered footnotes. One numbered list, inside the draft, of everything Likhat could not settle. The do-not-file line in the body. And an audit file holding the matter as received, the raw draft, the draft that shipped, each check's record and the chat log.
What none of this fixes
Here is the part that matters most, and it is short.
Our records hold a run where the first stage listed every fact a divorce petition needs. The drafting stage then made all of them up. The law it cited was correct. The client did not exist. Every check passed the draft, because the checks look for made-up law and there was none.
Code that checks citations does not check facts. Nothing in this design can confirm that the matter in the draft is the client's matter. That is why the warning line is in the body of the document. The whole run is written up in the day the pipeline wrote a perfect petition about facts that did not exist.
So, is it ChatGPT?
No, and the useful version of the answer is longer than one word.
One stage of the run uses a general model, the same kind of technology behind the chatbot a partner may have tried. If that were all there was, the mistakes in that chatbot would be our mistakes too.
The difference is everything around it. Likhat decides the document. Likhat finds the law and puts it in front of the writing. Likhat checks what comes back against the same files, strips what does not match, and lists what is still open. Likhat keeps the record.
Then a lawyer reads it. A lawyer reviews, a lawyer signs.
Likhat AI
Hardflip Technologies
11 September 2026 · २०८३ भदौ २६
Product facts in this piece are as of 11 September 2026.
This tool does not give legal advice. Every draft is usable only after review by a licensed advocate.