NiryatHub

Overview

Six questions, six offices, and no single answer

A farmer in Ilam has two tonnes of large cardamom. A co-operative in Bhaktapur has 500 hand-knotted carpets. Both know their product perfectly. Neither can tell you which country pays most for it, what that country charges at the border, who over there buys at their volume, what paperwork clears it, which crossing to use, or what the invoice should say. Today those are six different offices and a great deal of guessing. NiryatHub answers all six in one run, and shows its working at each step.

Built for Hackfest 2026 at MBM Campus, with team 4 NF.
The run starts with a spoken sentence and ends with a filed commercial invoice.
Every figure on screen carries a tag saying where it came from.
It runs with no database, no API key and no internet, because the data ships inside the app.

Problem

The margin sits in the questions nobody can answer

The reason this is worth building is arithmetic rather than sentiment. A duty of 8% on a $45,000 consignment is $3,600, which is frequently the whole margin, and the duty differs by destination on the same product. A German importer wanting eight-tonne lots is useless to somebody holding 200 kg. A landlocked country pays a different price at every crossing. Each of those is a knowable number that currently gets guessed.

Export prices differ by a factor of two between destinations, and move seasonally.
The paperwork list differs by destination and by product category, not only by country.
Nepal is landlocked. Birgunj is the sea corridor, and Kathmandu airport is four times the price.
None of this is secret. It is scattered across six institutions, which is not the same thing.

What was built

Eight screens, one run

Next.js 15 and React 19 on TypeScript, deployed on Vercel. Consignment, research, market, buyer, compliance, route, shipping and invoice, plus a dashboard of every order the browser has run, and a sources page that shows how the score is assembled. Every screen fits one viewport on a desktop, because this gets demonstrated from a stage and scrolling to find the next button loses the room.

The whole run's state lives in the visitor's own browser. No account, no server, nothing to leak.
One file, the repository seam, knows the trade figures are bundled rather than live.
Voice input is held open until the speaker taps stop, rather than trusting the browser's own silence timeout.
Below 720px the fixed frame lets go and the page scrolls, because a phone's chrome takes height the design assumed it had.

Diagram

Scoring

A score you can take apart

Ten destinations are scored 0-100 from seven components, each normalised against a fixed band and multiplied by a fixed weight. Cardamom into Germany scores 87, and the page will show you the seven lines that make it 87. The weights sum to exactly one, and a test fails if somebody retunes them until they do not.

Demand 25%, growth 20%, price 15%, market access 15%, competition 10%, logistics 10%, compliance 5%.
A missing figure is never a zero. The market scores null, reports insufficient data, and ranks last with its rank intact.
Anchors are fixed rather than relative, so adding an eleventh market cannot move the first ten.
Ties break on market id, so the same seed always produces the same board. The demo on stage is the one that was rehearsed.

Logistics

Pricing a landlocked country honestly

The route planner's real work is refusing to stop at the frontier. A planner that costs the run to the border sends everything by air, because Kathmandu is close. Every exit therefore carries a multiplier for the onward journey to the buyer: Birgunj 1.0, Biratnagar 1.08, Kakarbhitta 1.12, Bhairahawa 1.18, Rupaidiha 1.3, and Kathmandu airport 3.4.

The answer is not the obvious one. Out of Ilam, Biratnagar beats Birgunj by about $10 a tonne.
Hours are counted twice: to the border, and door to door. Only the second is comparable.
Routes are drawn along the Mahendra Highway, because the straight line from Ilam to Birgunj crosses India.
Eleven recurring disruptions turn one transit figure into three: planned, plan for, and worst case.

Compliance

The paperwork, sorted by who is asking for it

Fifty-eight requirements, each one a gate, a document, a label rule or a check, and each naming the authority that enforces it with a link to its page. The useful part is the sorting: a rule comes from Nepal, from the destination, or from this product in that destination, and a producer needs to know which, because they are three different phone calls.

Nepal side: EXIM code, certificate of origin, commercial invoice.
Destination side: importer EORI, TARIC classification, packaging registration.
Product side: pesticide residue limits for plant foods, labelling rules for textiles.
A gate stops the shipment; the distinction from a document is enforced in the type rather than in prose.

AI

What the model is allowed to touch

Two server routes call a language model, and both return prose only. Each composes its text deterministically first, offers the model the figures already computed, and keeps whichever it ends up with. No key, a refusal, a timeout or no network all land on the same deterministic text, labelled DETERMINISTIC on screen. The app is fully usable with no API key at all.

The model never originates a number. It rewrites a reading; it does not compute one.
Every figure carries a tag, SRC, CALC, AI or DEMO, and AI never sits on a number.
The fallback is not an error path. It is the default, and the model is the enhancement.
A deliberate constraint for a trade tool: a hallucinated tariff is worse than no tariff.

Testing

Sixty-two tests, because a demo is a claim

The seeded board claims cardamom scores 87 into Germany. A test asserts it. If somebody retunes a weight or edits a price, the suite fails before the stage does. The tests are aimed at the things that would embarrass the prototype in front of a judge rather than at coverage for its own sake.

Weights sum to exactly 1, scores stay inside 0-100, and missing data never becomes zero.
The seeded board still produces the figures the demo claims, and every export price beats the farm gate.
An eastern consignment prefers an eastern crossing on total cost, and door to door always exceeds the run to the border.
Every product's origin district reaches every modelled market, so a district with no leg to carry it cannot pass silently.

Outcome

What is real, and what is seeded

This is the part I would rather say first than be asked. All trade figures, meaning import values, growth, prices, shares and duties, are seeded. They mirror the structure of what UN Comtrade, ITC Trade Map and EU Access2Markets publish, and they are internally consistent, but they are not live extracts and must not be quoted as statistics. The compliance rules, the authorities, the links, the districts, the hubs and the border coordinates are real.

Every seeded figure is tagged DEMO on screen rather than mixed in silently.
Two importers per market are modelled, not the whole book.
Road distance is great-circle × 1.3, labelled an estimate, and nothing is priced from it.
Disruptions are recurring seasonal patterns, not a live incident feed. Nothing polls a road.

Reflection

What a deadline teaches that a semester does not

The VAT ledger taught me to put invariants in the database. Travora taught me that a system nobody uses has no invariants worth enforcing. NiryatHub taught the third thing: under a deadline the temptation is to make the demo look complete, and the discipline is to make it visibly incomplete in the right places. The tags on the numbers cost time we did not have, and they are the reason the prototype can be shown to somebody who actually works in trade.

Label the seeded data rather than hoping nobody asks. Somebody always asks.
Compute the deterministic answer first, then let the model rewrite it. Never the reverse.
Show the intermediate figures during the scoring pass instead of a spinner.
A prototype that admits its edges beats one that invents a match in front of a judge.

What's next

What we have not solved

The parts we did not build are the parts that make export hard. There is no financing, no liability model, and nothing here survives contact with a real field operation. Live trade data is a week behind the repository seam; the rest is not.

Live trade data behind the repository seam, replacing the seed without a component noticing.
Financing and payment risk: letters of credit are named in the data and modelled nowhere.
Liability, insurance, and who carries the loss when a corridor closes.
A second language. Every product, market and district already carries its Nepali name.