Sebastián Galčík · Slovakia · Central European hours

I build the delivery half.

You have already sold the automation. I am the person who builds it, on your terms, under your name. Your client never learns I exist.

Production systems built for GayaHakomAluprintMountberryFamisMTEC

What I have actually shipped

The first four cards are client work, in production. The rest are my own projects. Client names above are used with their permission, and references are available on request.

Banking & finance integration

Statement ingestion against three Slovak banks

Full OAuth token lifecycle with unattended refresh, CAMT053 statement parsing into a normalised SQL schema, and incremental daily runs with duplicate suppression scoped to the statement header — so a re-run mid-day cannot double-post a transaction. Deployed as a compiled Windows binary and as a Raspberry Pi service.

PSD2 / open bankingOAuth refreshCAMT053MSSQLbysquare QR

3 banks · production · runs daily, unattended

Customer portal

A customer portal in production, with online payments

PIN login by SMS or e-mail, invoices as PDF, contracts, service and return requests, and payments through TatraPay and CardPay with pay-by-square QR codes, in a web and a mobile version.

PHPMySQLJavaScriptTatraPayCardPaybysquare QR

live for a client · PHP and MySQL · payments end to end

Operational data plumbing

POS, production and IoT data into one database

Point-of-sale APIs reconciled into MSSQL on a schedule. Power-meter telemetry from IoT relays logged continuously. CNC programs parsed for cut length and part counts, part lists from Inventor and nesting CSVs from RADAN imported for production preparation. Excel and CSV workbooks loaded into the database and exported back by order number. Over a dozen of these, for real businesses that depend on them.

PythonpandasSQLAlchemypyodbcREST APIsCNC programsRaspberry Pi

12+ delivered systems · manufacturing, retail, facilities, agriculture

Operations automation

The small jobs that keep a business running

Backup checks with an e-mail alarm when something is missing, rule-based clean-up of old backups, scheduled FTP copies, and Raspberry Pi kiosks for attendance (RFID cards) and QR payments matched through the bank API. Packaged as Windows executables and run by Task Scheduler.

PythonPyInstallerTask SchedulerRaspberry PiFTP

unattended · Windows and Raspberry Pi · alarms by e-mail

Research infrastructure

A quantitative research programme, run like engineering

My own long-running project, and the reason I am worth trusting with something unattended. Automated backtesting over minute-resolution market data, orchestrated across parallel agents and four independent research lanes. Over 560 documented research rounds, each with a prediction registered before the test, and a graveyard registry that blocks re-testing an idea already killed. Method and strategy stay private; the discipline is the point.

agent orchestrationparquetMQL5falsification protocol

560+ rounds logged · 4 research lanes · predictions on record

How I build with AI agents

Codex CLI and Claude Code in one repository

Codex (codex exec) writes specifications, code and reviews. Claude Code runs the tests, judges the result and commits. The only channel between them is numbered message files in git, so every hand-over is on record. In six days the pair produced 56 Python files, 1,009 tests and 106 commits.

Codex CLIClaude Codegitagent protocoltests

106 commits · 1,009 tests · 6 days

A product of my own, live

Tender vopred: public contracts that are about to expire

Built on the open data of the Slovak Central Register of Contracts. It shows suppliers the public contracts that expire in one to twelve months. A Python pipeline, a static site on Cloudflare and scheduled GitHub Actions workflows, written with AI agents that I direct and check.

PythonCloudflareGitHub Actionsopen data

tendervopred.com · live · built with AI agents

Built in one evening, to show the pace

End-to-end short-form video pipeline

Writes the script, synthesises the voiceover, drives a real browser through a SaaS product (since retired) and records it with a synthetic cursor, times captions from the measured audio rather than from speech recognition, renders to 9:16 and publishes to three platforms.

PlaywrightFFmpegGemini APIInstagram GraphCloudflare Workers

9 modules · 3 platforms · 1 evening

A theme across all of it: the hard part is never the API call. It is the failure modes that report success — a caption field that silently discards programmatic input, a vendor 200 that is a job id rather than a published post, a recording with no cursor in frame, a backup nobody has ever restored. Those decide whether an automation survives a month unattended, and they are what I spend my time on.

How a first project works

  1. Send me one workflow you have already sold and not started.
  2. I quote a fixed price before writing anything, and agree the scope in writing.
  3. You get the first working version in five working days. You keep it either way.

No retainer, no minimum, no discovery phase to pay for. If the first build is good you send the next one. If it is not, you have lost a week you were not going to spend anyway.

Stack

I build with AI agents in Claude Code every day (1,436 sessions over 83 active days) and check their work with independent reviewers and tests, including mutation tests that catch deliberately inserted bugs.

Slovak rates, your timezone. I invoice as a registered EU sole trader, so there is no offshore payment friction and no overnight gap to manage.

Send me something to build

Reply to the email this link came from, or write to [email protected].

Fastest way to judge me: send the messiest spec you have. I will tell you what I would cut and what it costs, before you commit to anything.

Businesses can also hire me directly: routinebridge.com.