Projects

The problem: Demand is seasonal, the record of it already exists, and nobody can read it in time. A distributor watches an antiparasitic line spike in the same weeks it spiked last year — and learns about it by running out: the shelf is empty, the supplier's lead time is six weeks, and every order a veterinary clinic places that month goes to a competitor. That sale was not lost on price or on service. It was lost to a jump sitting unread in last year's figures while the reorder was decided from memory. Upstream, the manufacturer makes the mirror-image mistake, committing a production run to the lines that moved last quarter rather than the ones the coming season will ask for. Every industry with a seasonal curve and a lead time in front of it carries this gap; pharmacological distribution is simply where it costs the most, because the shortage is counted in prescriptions that cannot be filled.

A web application that turns a company's own sales and order history into a demand signal it can plan against — at SKU level, week by week, against the same weeks of previous years. It serves both sides of the same shelf: the manufacturer scheduling a production run, and the distributor raising a purchase order that has to clear a supplier's lead time before the season arrives. Veterinary pharmacological lines are the clearest case — antiparasitics, antibiotics and vaccines move on a seasonal curve set by herd cycles, husbandry calendars and weather, and the decision to make them or buy them is taken months before the demand appears.

Under the hood: The demand signal is arithmetic on the company's own transactions before it is anything cleverer, and every figure opens onto the orders behind it — a number nobody can trace is a number nobody acts on. Aggregation runs in PostgreSQL over indexed, partitioned history rather than in the application, and the scheduled jobs that rebuild each report run there too. Calendars are handled properly: retail weeks, fiscal quarters and a moving Thanksgiving do not line up with calendar dates, and a year-on-year comparison that drifts by a few days is worse than none at all. Where a curve is genuinely too noisy to read — a new line, a broken series, a one-off bulk order distorting a week — a model ranks what deserves a planner's attention. It does not replace the arithmetic, and it does not place orders.

The problem: Shops running a storefront, a stock spreadsheet and an accounts package that never agreed with each other — so every month began by reconciling three versions of the same week.

An end-to-end retail system: the storefront a customer buys from, the order as it moves through picking, packing and delivery, the inventory it draws down, and the customer record that remembers all of it. The owner gets margin and stock; the support team gets open orders and history without touching the owner's numbers.

Under the hood: PostgreSQL with row-level security, so the support role physically cannot read margin. Deployed on managed cloud with automated backups. The data model is configurable per industry — a furniture workshop and a pharmacy carry different attributes without forking the codebase.

The problem: Persian speakers type slowly on systems designed around Latin input, and existing dictation returns text that needs more correction than typing would have taken.

A native macOS application for writing Persian by speaking it, built for Persian first rather than added to a system designed around Latin input. The difference shows up in the details nobody else handles: half-spaces, Persian digits, and punctuation that faces the right way.

Under the hood: A native application rather than a browser tab, so it responds immediately and works everywhere on the machine. Built with an explicit boundary on where audio goes and what leaves the device.

The problem: Programmes written on paper or in chat messages, lost between sessions, with no record of what a student actually did or how they progressed.

Coaches assign exercises to each student and record what was done, so a student arriving at the gym knows their programme without asking. Body measurements load into the athlete's record, turning progress into a history rather than a memory.

Under the hood: Cloud-backed so the coach's phone on the floor and the office computer show the same data. Designed for the small screen it will mostly be used on.