01
AI and data
Put intelligence behind the work still done by hand, and keep people in charge of the decisions.
Software company · Lahore, Pakistan
We build the systems companies run on.
ERP and internal platforms, products made to sell, and the AI that goes inside both. Designed, built and looked after by one team of more than fifty people.
See what we built for
Orders in one file.
Ad spend in another.
Stock in someone’s head.
When the numbers disagree, nobody knows which one to trust.
We turn all of it
into one system.
What we do
Five practices.
One team accountable
for all of it.
Tell us the problem in your own words. We’ll tell you plainly which part of it is ours.
02
Commerce
Shopify at the point where one store has become several and the spreadsheets can’t keep up.
03
Product engineering
The software itself: the systems a company runs on, the products it sells, and the apps its people use.
04
Cloud and platform
Infrastructure that is repeatable, observable and safe to change, with a rehearsed way back.
05
Working with us
Long-term care for live systems, or engineers embedded directly in your own team.
Selected work
Three clients.
Three very different problems.
Choose a name to see what we built. Each of them agreed to be named here, and you’re welcome to ask them about the work.
Illustration with sample figures
Dubai · Multi-store Shopify
One system over five sources that disagreed.
Jansher Khan sells on Shopify out of Dubai. Every weekly number came from five places that didn’t talk to each other: Shopify orders, Meta and TikTok ad spend, a courier manifest and a payment settlement file. Column names, date formats and currencies all differed, so finding out what a week had really earned meant reconciling them by hand.
- React
- TypeScript
- Supabase
- Postgres
What we built
- Each file is recognised by its own column headers, normalised into one shape and deduplicated by order reference, so the same export can’t inflate a total.
- A written rule decides which source wins when two of them disagree.
- Screens for orders, ads, couriers, settlements, returns, inventory and staff attendance, each with its own filters and export.
- Profit and loss computed from the normalised data whenever someone asks for it.
- Access scoped per store and per role from the first week.
What changed
It is one system now: several Shopify stores, the accounts behind them, inventory, settlements and the people who run them. “What did we actually make?” is answered by opening a page instead of five files, and the whole team works inside the system instead of one person exporting out of it.
Services behind this work
From the public demo at pulseboard.com.pk
United States · Pulseboard
Pulseboard: every sale, live on the office wall.
Haseeb Iqbal wanted a product he could sell to people running several Shopify stores, people who already owned dashboards and never opened them. So we built a display rather than a dashboard.
- Next.js
- Three.js
- Tone.js
- Neon Postgres
- Upstash Redis
What we built
- Every sale, from up to fifteen stores, fires a beam of light and a musical note on a 3D map of Pakistan. The bigger the order, the taller the beam.
- Orders arrive through Shopify’s Admin API and webhooks along one write path, and reach the screen over server-sent events, built to run all day untouched.
- Scroll and the scene rewinds into hand-drawn charts: revenue against orders, top cities, the cash-on-delivery funnel and the hours when the country buys.
- A TV pairs by holding a phone up to a code on the screen, with no typing on a remote.
- Only a signed webhook from the payment provider can change an account’s plan.
What changed
Pulseboard is live at pulseboard.com.pk on three plans priced by how many stores a customer runs, each with a free trial, and its billing has been running since October 2026. It sells under its own name, not ours.
Services behind this work
Review queue3 waiting
Suggested reorder
Reorder 240 units of item A-1042
Sales over the last 60 days leave 31 days of cover. The supplier takes 38.
Suspected duplicate
“Steel bolt M8 x 40” and “M8x40 bolt, steel”
Two suppliers, very likely one product. Merge them only if you agree.
Supplier document
Invoice 7731: 3 fields need a look
Read from a supplier’s PDF. It reaches stock only after a person checks it.
Queue clear
Every decision has a name on it.
That is the rule on this system: the model reads and estimates, and a person signs.
Try it. Illustration with sample data
AI-assisted ERP · Run by us
AI-assisted inventory, with a person signing every decision.
Entrepreneur Hub asked us to run their ERP, not just build one and leave. Inventory sits at the centre of it, and five jobs were being done from memory and a spreadsheet: when to reorder, whether two suppliers ship the same product under different names, retyping supplier documents, tracking payments apart from stock, and keeping a team pointed at all of it.
- Python
- FastAPI
- Postgres
- pgvector
- Next.js
What we built
- Reorder forecasts run overnight and arrive as a suggestion with the numbers behind it, never as an order.
- Product matching uses embeddings kept in the same database as the transactions. A suspected duplicate appears as a pair for someone to confirm.
- Supplier documents are extracted into fields and land on a review screen before they reach stock.
- Payments sit beside the inventory they paid for, and responsibilities live inside the system.
- All five jobs end at a screen where a person approves, rejects or corrects.
What changed
It is running, and we run it. A reorder now has its working attached, a merge is someone’s decision that can be traced back, and a supplier’s paperwork is a set of checked fields rather than a retyped page.
Services behind this work
AI practice
AI that shows its working.
Plenty of companies have added AI to a price list. For us it is a practice, built on three commitments you can hold us to.
Answers that cite their sources
Retrieval shows what each answer was drawn from, so the person acting on it can check it for themselves.
Measured before every release
An evaluation set runs before a change ships, so a new prompt is a measured change rather than a hopeful one.
A cost you can read
Cost and latency per request are requirements from the first week, not a surprise on the first invoice.
The decision we write down first
What may the model decide, and what may it only suggest? We agree that with you in plain words before anything is built, then build the screen where a person approves, rejects or corrects.
How we work
You can open it in the first week.
- 01
Scope
We write down what the system has to do before anyone debates how. You get a document you can challenge while changes are still cheap.
- 02
Build in the open
Working software from week one, deployed where you can open it on your phone. You see progress instead of hearing about it.
- 03
Harden
Error paths, empty states, slow networks, screen readers, the largest screens and the smallest. This is the distance between a demo and a product.
- 04
Hand over
The repository, the pipeline, the runbook and the reasoning behind the hard decisions. Your team can carry on without us.
Technology
The tools we stand behind.
We name our stack in public, so you know what you’re getting before the first call. Python and Postgres sit at the centre of our AI and data work, TypeScript and React across our product work, and Rust wherever a service has to be fast and can’t afford to fall over.
- Rust
- Python
- TypeScript
- React
- Next.js
- Node.js
- React Native
- Flutter
- Postgres
- pgvector
- FastAPI
- LLMs
- Shopify
- AWS
- GCP
- Docker
- Supabase
- Redis
- Three.js
- Tone.js
About
A studio first.
A company since 2025.
We spent years as a studio, building for international clients and growing our team around that work. We incorporated in 2025, which is why a young company has more than fifty people. The company is a year old. The people in it are not.
We build three kinds of thing: internal systems, which are most of our work and the least visible part of it; products made to sell, one of which is live and taking money under its owner’s name; and the AI that goes inside both.
Every engagement ends with a handover we’re willing to be judged on. If we disappeared, your team should be able to keep shipping.
- Lahore
- Pakistan, working with international clients
- 2025
- Incorporated, after years as a studio
- 50+
- People on the team
Careers
Send us your work, not your CV.
If you would rather build systems like these than commission them, we want to see what you’ve made.
Data engineer Own the path from raw exports to the numbers a business runs on. Correctness is the job. See the details
- Recognise files by their headers, normalise them, and deduplicate by order reference.
- Write down which source wins when two disagree, and keep the tests that prove last month still reconciles.
- Treat Postgres as the centre of gravity: query plans, indexes and constraints, with Python for the pipelines.
- Work beside models: embeddings for product matching, document extraction, and forecasts a person approves. You don’t need to have trained one.
One position, on site with the team. We discuss pay in the first conversation.
Apply for this roleNot a data engineer?
Send us the work anyway: a repository, something you shipped, something hard you fixed. Tell us which part of the work you want to be in.
Contact
Tell us what you’re building.
Two or three sentences is enough: what it has to do, the part you already know is hard, and when it needs to be live. You’ll get questions back, not a sales deck.
If you can’t describe it in an email, say so, and we’ll sign your NDA before the first call.