Nepal's tax & audit, answered with the law.
An AI legal and tax intelligence platform that answers Nepal's tax and law questions with the exact statutory section and court ruling - built for CAs, auditors, lawyers, businesses, and foreign investors.
Nepal's tax and legal framework is vast, layered, and slow to search by hand. The relevant answer to a single question might live across the Finance Act, the Income Tax Act, a VAT circular, an NRB directive, and a Supreme Court ruling - each in a different document, often only in print or scattered PDFs. Chartered accountants, auditors, and lawyers were burning hours hunting for the precise section and precedent, and businesses and foreign investors had almost no way to get an authoritative, cited answer without paying for a consult. Professionals needed answers in seconds, grounded in the actual law, not a confident guess.
We built Vidhica as an AI legal and tax research engine grounded in 184+ indexed Nepali laws - the Finance Act 2082, Income Tax Act 2058, VAT Act 2052, Supreme Court rulings, and NRB directives among them. Every answer cites the exact section and the relevant precedent, so professionals can trust and verify it instantly. On top of research we shipped an OCR-powered invoice vouching tool, auto-drafted audit reports, a past-exam answer bank for CA students, and an ICAN-verified CA directory where users can book consults starting from NPR 1,500. Two model providers - OpenAI and Anthropic - back the reasoning layer so the platform can route each query to the model best suited to it.
Reading the law before writing the code
Tax and audit answers are worthless without a citation, so the first phase was not technical. We sat with practising chartered accountants, collected the questions they actually get asked, and built the evaluation set that everything after was measured against.
Retrieval first, generation last
The hard part of a legal AI product is fetching the right passage, not phrasing the answer. We invested in chunking the corpus along statutory boundaries and in metadata that lets a query filter by act, section and year, then treated the model as the thin layer on top.
Grounded, cited, and measured
Every answer carries its source passages, and we scored each release against the evaluation set before shipping it. When accuracy regressed we could point at which retrieval change caused it rather than guessing at the prompt.
Cited AI research
Ask any tax or law question in plain language and get an answer grounded in the exact statutory section and court ruling, drawn from 184+ indexed Nepali laws and directives.
Living legal corpus
An indexed library spanning the Finance Act 2082, Income Tax Act 2058, VAT Act 2052, Supreme Court rulings, and NRB directives, kept current as the law changes.
OCR invoice vouching
Upload invoices and the OCR pipeline extracts, structures, and cross-checks the data, turning a manual audit chore into a few seconds of review.
Auto-drafted audit reports
Generate first-draft audit reports from structured inputs, giving auditors a defensible starting point instead of a blank page.
CA exam answer bank
A searchable bank of past exam answers helps CA students study against real questions with model responses.
ICAN-verified directory
A directory of ICAN-verified chartered accountants lets users book consults starting from NPR 1,500 when they need a human in the loop.
Dual-model reasoning
OpenAI and Anthropic models sit behind a routing layer, so each query is handled by the model best suited to its complexity and length.
Project FAQ
About this
build
Common questions about how we approached and shipped Vidhica.
Every answer is grounded in the indexed corpus and returns the specific section and ruling it relied on. Because the citation is shown alongside the answer, professionals can verify the source directly rather than trusting an unsourced response.
Let's ship something like this.
Vidhica went from concept to production in 4 months. Tell us what you're building and we'll scope exactly how we'd ship it.