Software that outlives its MVP.

The operating system an agency runs on, with AI built into the data rather than bolted beside it. CRM through to cash — accounts, deals, proposals, contracts — then native projects with work items, cycles and worklogs, HRM and payroll records, a multi-currency finance module built on a service catalogue, plus LMS, recruitment, OKRs, a wiki and a client portal. Twenty-five specialist agents across sales, delivery, engineering, marketing, talent and internal ops work over the same organisation-scoped store, so an agent answers from the agency's own accounts, projects and documents instead of from a chat window that knows nothing about them.

What was technically hard

Twenty-five agents reading one knowledge base, none of them able to see across a tenant boundary. Retrieval is scoped at the query layer — the organisation is an argument to the search, not a filter applied to results afterwards — and the isolation is enforced again underneath by Postgres row-level security, because a filter you forget is a data leak and the second line has to hold when the first is wrong. Money made it harder still: every amount is a value object carrying its currency and the FX rate captured at transaction time, so a figure stays true to the day it was recorded rather than to today's rate.

An AI takeoff platform for construction estimating. It reads a PDF floor plan and returns structured quantities — rooms with areas and perimeters, openings with widths and heights, dimension chains and building sections — which an estimator exports as CSV or JSON and bids from, instead of measuring the drawing by hand.

What was technically hard

A language model returns plausible JSON, not a schema. Plans name their levels inconsistently, quote dimensions in whatever the drafter used, and omit fields entirely, so nothing from the model is trusted directly: every number is coerced before it becomes a row, levels are derived from the floor map rather than read off the labels, and every room and opening carries its own confidence score. An estimator prices a bid off these figures, so a wrong number that looks certain is far worse than one marked uncertain.

A multi-tenant school platform where any teacher or school gets an isolated workspace of their own. Lesson planning with media, tests and question banks, homework with media submissions, attendance, a gradebook, class projects with teams, messaging and announcements, and a shared library — with parents given read access to their own children's progress and a line to the teacher. A single tutor running a school of one is a first-class case rather than a workaround, and one account can hold different roles in different schools. English and Ukrainian throughout, with AI help for building lesson content over an MCP server.

What was technically hard

A gradebook has to be honest about a term that is only half finished. The naive version divides by everything planned, so a pupil three weeks into a course reads as failing, and a parent who sees that number acts on it. Totals here accumulate only the items a pupil actually has a grade for; attendance is measured against lessons that were actually marked, so an unrecorded lesson never counts against anyone and an excused absence never punishes. The other half was making it fast: every gradebook view assembles from a fixed number of queries no matter how many pupils are in the class, and a test guards that count so a later change cannot quietly reintroduce a per-pupil lookup.