We make complexity dependable.

+91 93117 30107

All work

Steel Pro

Every bar and every cut piece, tracked from the order to the dispatch gate.

Built for
A steel bar cutting, machining and trading company in Faridabad
Sector
Steel cutting and machining
Runs on
iPhone, Android, web
Built on
React Native, PostgreSQL, nine server functions
Status
Stock live on the floor

What it does

A steel service centre runs on exceptions. An order arrives as an Excel sheet on WhatsApp. Its sizes have to be standardised before anything is cut. A bar comes up short, a cut piece weighs in under, a machined part comes back from an outside vendor with a step still to do. The plant was handling all of it through WhatsApp, Excel and the shop-floor manager's memory.

Steel Pro carries the whole plant on one system: seventeen staff roles across two plots, each with its own screens, all writing to one database. It reads incoming order sheets with AI and has a person confirm every row, standardises sizes against the plant's own catalogue, queues work for three saws, weighs every piece, runs the machining plot and its outside vendors, prints labels, bills and dispatch papers on the plant's own printers, and gives the owner an assistant that answers questions about the business by voice.

Stock tracking is live on the factory floor. The chain from order intake to dispatch is built end to end, and each stage is being proven with the plant's staff before the owner switches it on.

17

staff roles, each with its own screens

2

plots, run as one factory

3

saws with their own live queues

5

ways an order can arrive, from WhatsApp to Excel

1,900

pieces in the largest single order so far

9

server functions behind the phones

561

database revisions, each one reviewed

295

automated tests on the plant's own rules

Before

  • Almost every order arrived as an Excel sheet over WhatsApp, and a few as photographs of paper.
  • Stock was tracked by hand through the shop-floor manager.
  • Bills were written up separately in the billing software.
  • The sales calling sheet was a separate Excel workbook.
  • Field-visit distances were argued over.

After

  • Software reads each order sheet, and a person confirms every row before it moves.
  • Sizes are standardised against the plant's own catalogue, and the customer can confirm them from a link.
  • Every bar and every cut piece carries a printed QR label.
  • A short bar reaches Purchase the moment the floor says so.
  • A photograph of a bill supplies the amount.
  • Road distance for each visit is measured by Google Maps.

An order, end to end.

  1. Arrives

    By WhatsApp share, camera, gallery, PDF or Excel. Each file becomes one import that can be sent again without creating a second order.

  2. Read

    AI reads the sheet in blocks and matches every answer back to the sheet's own lines, so no row goes missing quietly.

  3. Confirmed

    A person checks every row. Reading an order is never the same as approving it.

  4. Standardised

    One shared engine maps each asked size to a size the plant can cut. The customer can confirm the choice from a link.

  5. Cut

    Each saw sees only work it can do. The operator scans a matching bar, and every cut is recorded piece by piece.

  6. Weighed

    Every piece is OK or not good, and not good must say over or under. Under needs exact measurements and a photograph.

  7. Machined

    Four stations, from milling to grinding. Work sent to outside vendors comes back booked for only the steps actually done.

  8. Billed

    Ordinary and machining bills from one screen, drawing on one shared book of bill numbers.

  9. Dispatched

    A dispatch sheet or an inspection report leaves with the steel, and a board on the office television shows what is going out.

How it is built.

Where people work

  • Phone app

    iPhone and Android. Each person signs in with a one-time code to their own phone and works in one of seventeen roles.

  • Office web console

    Quotations, users, dispatch promises and office decisions, on the same live data.

  • Customer link

    The only page without a login. A customer picks from fixed size options, and nothing reaches the order until the office accepts it.

  • Dispatch board

    A read-only board for the office television.

Where the rules live

  • One database

    PostgreSQL with row-level security: the database itself decides who may read or change each row. A screen is never the security.

  • Server functions

    Nine, each with one job: reading orders and bills, the sales inbox, quotation follow-ups, road distances, push alerts and the owner's assistant.

  • Live updates

    A cut on the floor shows on the office screen within moments, and every screen re-reads on its own when it reconnects.

What it connects to

  • AI document reading

    Order sheets, photographs and supplier bills.

  • Gmail

    The sales team's shared inbox, with quotations attached from the same document the office prints.

  • Google Maps

    Measured road distance for every field visit.

  • The plant's printers

    A thermal label printer and two A4 printers on the plant's own network.

The hard parts.

An order sheet is read in blocks, then checked line by line

A long Excel order never goes to the AI whole: past about seventy rows a model can stop early and still return a confident answer. Steel Pro cuts each sheet into blocks that carry their own headings, matches every answer back to the sheet's lines, asks again for anything unread, and stops loudly after three tries rather than save half an order.

Under the hoodChunked extraction, line-level reconciliation, and runs that pick up where they stopped before a time limit cuts them off.

Every list comes back whole

The database interface quietly returns at most 1,000 rows per answer and raises no error. Here every list is read in pages, a checker scans the whole code base for any read that is not, and a tripwire turns an answer of exactly 1,000 rows into a visible error, so a screen is never quietly incomplete.

Under the hoodPaging ordered on unique keys, static analysis in the release checks, a guard on every request.

A retry can never count twice

Factory Wi-Fi drops. Every action that can be repeated carries an id made on the phone, so a second tap or a resent request finds the first one instead of making a duplicate. Related rows change together in one transaction, and every cut, weighing and scan is kept in an execution history.

Under the hoodIdempotency keys, claim-before-work, guarded timestamps, transactional batch writes.

Live, but never trusted blindly

Screens change the moment something happens on the floor, but a live update is treated as a hint, not the truth. Every screen re-reads when it reconnects, bursts of changes merge into one read, late answers are thrown away, and a screen that is not in front listens to nothing, so phones stay quick on a busy day.

Under the hoodRealtime subscriptions, debounced refetching, request-generation counters, channels tied to focus.

The assistant can look, never touch

The owner asks questions about the business by voice. The assistant answers by writing database queries, and those run as a read-only database user with a statement guard, a time limit and a cap on rows. Its instructions are not what keeps the data safe; the database is.

Under the hoodNatural language to SQL, a read-only role that cannot log in, a statement guard, timeout and row cap.

Scrap, measured machine by machine

Scrap is worked out from each piece's received size and its machined size, in kilograms at the grade's density, and split between the machines by which one cut first. An outside vendor's cuts are counted apart. The owner sees a scrap rate for every machine, every month.

Under the hoodPer-piece volume accounting, machine order from recorded milestones, vendor attribution.

Every desk, scored every week

Each desk has its own deadlines: a shortage ordered within an hour, a piece cut before five weighed the same day. Misses are counted from the work itself, each working day a thing stays undone counts again, and every desk's score is ready for the Monday meeting.

Under the hoodDeadline rules evaluated inside the database against a working-hours calendar.

Printing on the plant's own printers

Labels go to a thermal label printer as raw printer commands over Wi-Fi. A4 sheets go straight to an HP and an Epson, each on its own proven path, with the phone's print dialog as the fallback. Preview, print and PDF share one renderer, so what is checked on screen is what comes out of the printer.

Under the hoodTSPL over TCP, IPP and IPPS print paths, a single A4 document renderer.

Department by department.

Stock

  • Bars loaded with QR labels printed before the bar is saved
  • Offcuts tracked back to the bar they came from
  • Length and weight reconciled against the grade's density
  • Re-measure checks that flag doubtful numbers for a person to confirm

Sales

  • Photograph an enquiry, price it, send a PDF quotation
  • Quotations followed up once they are delivered
  • Calling sheet with monthly targets
  • Shared inbox with follow-up tracking

Orders

  • Order sheets read from Excel, photos, PDFs or WhatsApp
  • Sizes standardised before cutting
  • Customers confirm sizes from a link
  • One firm is one customer, however its name is spelt

Purchase

  • Shortages arrive the moment the floor commits them
  • Gas-cut, standard and exception orders in their own lanes
  • Demand closes only when the steel is booked into stock

Saws and weighing

  • Operators scan a bar and cut
  • Every piece weighed and marked OK or not good
  • Shop-floor decisions recorded in Hindi
  • An under-size piece goes to the customer with its measurements and photograph

Machining

  • Four stations with their own queues, and runs that carry over a shift change
  • Work sent to outside vendors and booked back piece by piece
  • Scrap worked out for each machine

Billing and dispatch

  • Bills, challans and month-end bills
  • Printed dispatch, loading and inspection sheets
  • A dispatch board for the office television

Field sales

  • Visits logged with location, voice notes and visiting cards
  • Travel allowance from measured road distance

Owner

  • Stock, people, machine and rider reports
  • An assistant that answers questions about the business by voice
  • A weekly score for every desk
  • Scrap rate for each machine, month by month

How it was tested

  • 295 automated tests on the plant's own rules: sizes, weights, rates, routes and printed sheets.
  • 17 standing checks on the code itself, from the 1,000-row guard to the order-sheet splitter.
  • Every database change written as a reviewed migration, and applied to a separate test database as well as the live one.
  • Stock went live first. Each later stage runs with the plant's staff before the owner switches it on.