Work

We run this stack ourselves

Both systems below run our own venues. That is deliberate: we would rather show you something carrying real money and real staff every day than a pilot we talked a client into.

Our own taproom

Replacing a POS licence with a Telegram mini app

A venue paying a per-till licence for a commercial POS wanted off the subscription without going back to pen and paper. We rebuilt the till as a Telegram mini app.

In-house deployment — we run this venue ourselves.

Point of sale mini app interface

The problem

Monthly per-till cost, a tablet that had to survive a bar, and no way to check takings without opening a back office on a laptop.

What we did

  • Catalogue, pricing rules and staff records moved into Postgres as the single source of truth
  • The till rebuilt as a Telegram mini app against an action-routed API
  • A separate admin API with role checks resolved from the caller's Telegram account
  • Requests signature-verified inside Postgres, every API call logged
  • A nightly sales summary pushed to the manager's chat at 23:30

Where it landed

  • Every staff phone is a till; no dedicated hardware to buy or break
  • Takings visible on demand, and pushed nightly without anyone asking
  • Admin actions gated server-side, so staff access does not depend on UI hiding
  • No per-seat licence
The same venue

Twelve destinations, one approval step

Event promotion was not happening because posting to five apps by hand is a chore nobody picks up. We built a queue with a review step and let a schedule do the rest.

In-house deployment.

Social post approval queue interface

The problem

Events went unpromoted. Fully automatic posting was not acceptable — the venue's voice matters and an unreviewed post is a liability.

What we did

  • Queue held in a data table, one row per post, with an explicit approved flag
  • Telegram mini app to preview, edit, approve or deny each post
  • Scheduler sweeps every five minutes and publishes only approved, due posts
  • Fan-out to twelve destinations — including the venue's own TV screens and website blog — each reporting separately
  • AI flyer builder drawing on a weekly-indexed library of the venue's own photos

Where it landed

  • Nothing publishes without a human approving it
  • Posts with no photo still go out, with generated artwork
  • Per-destination failures surface immediately in Telegram instead of silently
  • Review takes under a minute from a phone
Also in production

Smaller systems, same approach

Customer service bot

A Telegram front desk that has been answering for the Automaton Labs shop since August 2025 — catalogue, orders and profile questions handled behind an authenticated boundary.

Catalogue shop mini app

The Automaton Labs storefront: catalogue, orders and customer profiles served as a Telegram mini app, with every request's identity verified before it is trusted.

Bilingual staff messaging

Type an instruction in English; the staff group receives it simplified, translated into Khmer, and read aloud as a voice note — because a message nobody understands is a message nobody follows.

Receipt capture and reconciliation

Photograph a receipt in Telegram and it is filed. Every three days the receipt folder is reconciled against the expense ledger: missing rows appended, duplicates and currency or date anomalies flagged, and a summary posted to the accounting group.

Tell us what keeps going wrong

A conversation costs nothing and often ends with us saying automation is not worth it for a given job. That is a useful answer too.