Ush

An AI meeting scheduler for executive assistants. Architect and sole developer, start to finish.

The problem

Executive assistants pick meeting slots using dozens of implicit factors: habits, travel times, how much a meeting matters, the other party’s constraints. That logic lives entirely in their heads and cannot be delegated or scaled.

There was nothing else to work from. No existing process to automate, no system to integrate. Just judgement in somebody’s head, which had to be made explicit and computable from scratch.

The insight

Scheduling is not a workflow to automate. It is a constrained optimisation problem wearing a workflow’s clothes. Once the assistant’s judgement is written down as weighted parameters, “pick a slot” stops being a sequence of steps and becomes a search for the provably best option.

The architecture

Hard constraints filter the candidates, weighted soft constraints rank what survives, and the language model only writes the human-readable explanation. The decision itself comes out of a transparent scoring engine, so every suggestion is debuggable and reproducible instead of a black box.

Around that engine: preferences learned from calendar history, calendars read and written in both directions, real-time travel time between consecutive meetings so physically impossible slots are rejected outright, and tenancy scoped per user by construction, so one assistant can schedule for several principals with nothing leaking between them. The system drafts the reply and never auto-sends it. A person is in the loop on every outbound message, by design and by the granted permission scopes.

Python, Flask and a separate background worker, PostgreSQL for everything including the job queue, Docker. Polling instead of push, one database instead of three services, self-hosted compute with cloud APIs only. Each of those is a cost and control trade-off, made deliberately and written down.

What it took

It became genuinely usable once travel-time checks and per-user tenancy landed. The stakeholders arrived holding UI prototypes they had already paid someone else to draw, and the shipped product replaced them, because it solved what they needed instead of what they had drawn.

Scheduling round-trips came down to a single interaction on average. That is the only outcome number on record, and it is the only one published here.

The client company was about three people, funded by one of its own founders — the same shape of buyer as the other work on this site.

What is not on this page

No client name, no logo, no quote. No usage numbers either: how many assistants or principals use it, and how many meetings it has handled, were never measured.