Oldrich Works

Software, AI agents, data and 3D · supervised by Oldrich Dvorak

Working results, shown to you before we ask to be paid.

Oldrich Works is a studio built around a specialised AI-agent pipeline, supervised by Oldrich Dvorak, who stands behind every result. The same pipeline runs our own e-commerce company every day — the shop, product feeds and ads, customer support and the servers — with real customers and real money. The case studies here come from that work; the names of the businesses behind them are withheld.

  • 30 000+commits across our own repositories
  • Over a millionlines of code
  • About six monthsin daily production on our own systems

Built and run with this pipeline: Hive, the agent platform that operates our companies (shop, ads, customer support, servers, code); Atlas, a multi-agent research platform with persistent memory; a research programme whose result is published with a DOI; and five open-source tools for AI coding agents.

How we work

The same loop for a one-hour fix and a month-long build. Larger work is split into milestones you can check one by one.

  1. Understand the goal

    We say back in one sentence what you want to achieve, and ask only the questions that change the plan or the price.

  2. Plan

    Scope, milestones with dates and a price per milestone — plus the risks you may not have seen, before any money moves.

  3. Build

    In your stack and conventions, in a separate secured folder for your project. Work continues overnight.

  4. Independent review

    A model from a different AI family reviews every change, looking for the blind spots of the one that wrote it.

  5. Tests

    Automated tests plus a run on the real target: a browser, a staging copy, your sample data.

  6. Hand-over

    You see it working first. Then the code, tests, run notes and hand-over documents. No lock-in.

Nothing paid to us up front
You see a working preview before we ask you to fund a milestone; the platform holds the money until you release it.
14 days of free fixes
After hand-over, anything that does not work as agreed is fixed at no charge.
Replies within minutes
Around the clock, every day. You wake up to progress, not to silence.
Your project stays on record
You never re-explain the context, even months later.

Why a second model family. On the same 15-bug test project, cross-family review caught 80–87 % of the bugs; the common pattern of five parallel reviewers from one family caught 40 %. Measured with full-review, our open-source review tool.

3D modelling, rendering and CAD

Product renders and turntables, interiors, game-ready assets and print-ready parts. Everything below is modelled, lit and rendered from scripts in Blender and the build123d CAD kernel, with CC0 textures and no AI image generators. These four pieces are our own demonstration work, not client commissions.

Second interior view: sofa wall with framed print, lamp and coffee-table props
Interior, second camera view
Wireframe render of the low-poly treasure chest showing its topology
Game asset, wireframe of the 892-triangle mesh
Frame from the product turntable: serum bottle and carton on a travertine plinth
Product, frame from the turntable

E-commerce and web engineering

Shopware, PrestaShop and custom web apps: storefront speed, plugins and modules, checkout and carrier integrations, conversion measurement, multilingual stores and fast sites on Cloudflare Workers — and the servers they run on. All from production work for businesses we run ourselves.

Shopware storefront: mobile load time cut from about 4 s to about 3 s, no rebuild needed

EU direct-to-consumer brand · Shopware 6 · 9 storefront languages

ProblemSlow mobile pages although the server answered in 50–170 ms. The open question was a costly headless rebuild.

DeliveredDiagnosed page weight as the cause, then shipped subset web fonts, modern prefetching with checkout excluded and a corrected hero-image preload. Lighthouse before and after, two independent reviews.

ResultMobile LCP ≈ 4 s → ≈ 2.9–3.1 s on home, category and product pages, live since 4 October 2026. The headless rebuild was not justified.

CS-01Shopware 6 · Twig · WOFF2 subsetting · Speculation Rules · Lighthouse · Varnish · Cloudflare

Content site moved off a heavy e-shop CMS: mobile Lighthouse 56–58 → 88–100

Central-European coaching practice · bilingual site · formerly one storefront of a PrestaShop multistore

ProblemAn audit of all 90 pages: main content appeared after about 10 s on mobile; pages weighed 1.5–1.85 MB.

DeliveredA static site generated from the existing pages and served from Cloudflare Workers: purged CSS, self-hosted fonts, WebP images, 74 redirects from the old domain, then a 48-run speed test after launch.

ResultMobile performance 56–58 → 88–100; LCP 9.5–10.9 s → 1.25–3.65 s; page weight −40 to −70 %. Live since 29 September 2026.

CS-02Cloudflare Workers · Node build · PurgeCSS · Wrangler · Lighthouse · Playwright

PrestaShop multistore: a core PHP notice firing 1 260 times in 48 hours brought to zero

Four storefronts on one PrestaShop 8.1.7 core and database

ProblemProduct URLs belonging to sibling shops threw a core notice and redirected to malformed URLs; clearing the cache ran out of memory.

DeliveredTwo upgrade-safe class overrides, a config drop-in for the template warmer, a lock-guarded activation with readback on every shop and notes for future core upgrades.

Result1 260 hits per 48 h → 0 in the following 24 h; output identical to core on 53 products; cache clear OK in 6 s at 512 MB.

CS-04PrestaShop 8.1.7 · PHP · class overrides · Symfony DI · MariaDB · Docker

Multi-market carrier checkout: home delivery repaired, 8,093 of 8,106 phone numbers valid

EU direct-to-consumer brand · Shopware 6 · pickup-point and home delivery in 8 countries

ProblemGuest checkouts failed home-delivery validation; carrier exports failed on malformed phone numbers.

DeliveredDelivery resolution that works before an address exists, one shared phone normaliser for both export paths and repaired parity tests. Independent review stopped one release for a calling-code defect, fixed before it landed.

ResultHome delivery in checkout repaired in two markets; in production 8,093 of 8,106 replayed phone numbers valid; on staging the vendor suite passed 162/162 across three seeds.

CS-06Shopware 6 plugin · PHP 8.4 · carrier API · PHPUnit · cross-model review

Server-side purchase tracking through Cloudflare Zaraz, deduplicated with the browser

EU direct-to-consumer brand · Shopware 6 · four currencies

ProblemPurchases had to reach GA4 and Meta from the server without double counting; order lines without a barcode silently dropped out.

DeliveredA Shopware plugin sending Purchase and Refund events through Zaraz with one shared event id for browser and server, a production revalidator and a barcode fallback that never fabricates a GTIN.

ResultBrowser/server deduplication proof passed; the fallback for order lines without a barcode has been live since 27 September 2026, with no fabricated GTIN.

CS-16Shopware 6 plugin · Cloudflare Zaraz · GA4 · Meta Conversions API · systemd timer

Comparison-shopping engine switched conversion measurement back on

Direct-to-consumer brand · Shopware 6 · national comparison-shopping engine

ProblemThe server uploader re-sent old orders whenever they changed; the engine counted them as orphaned conversions and turned measurement off.

DeliveredRoot cause measured in the conversion log; a fail-closed guard that sends only orders able to pair with a storefront visit, a 72-hour window taken from the measured payment lag, a dry-run mode, network-proof tests.

Result30-day production dry run: 0 of 498 paid orders would be re-sent; the window keeps 99.8 % of paid orders; measurement re-enabled on 30 September 2026.

CS-17Shopware 6 plugin · PHP 8.4 · conversion API · MariaDB · PHP test harness

Booking, QR prepayment and automatic invoices on Cloudflare Workers + D1

Solo consulting practice · paid 1:1 online sessions · two currencies

ProblemIntake and booking ran on third-party widgets. The owner wanted real calendar slots paid before the meeting, invoices and instant alerts — without trackers.

DeliveredA Worker API on D1 with Google Calendar free/busy, one booking per slot, payment QR codes for both currencies, bilingual PDF invoices, bot protection and a reminder job that sends exactly once.

ResultAPI tests 46/46 and mail tests 15/15 passing (each QR decodes to the exact payment string); independent review: 0 high findings; version 1.4.1 live, all health checks green.

CS-12Cloudflare Workers · D1 · Cron Triggers · Google Calendar API · payment QR · PDF · Turnstile

QR bank-transfer checkout, confirmed automatically through a bank API

Consumer web app · digital products · Czech market

ProblemNo card gateway at launch, yet paid content had to unlock by itself after a bank transfer — and never be given away by a retry or a tampered request.

DeliveredOrders with a unique payment reference; a sync job reads the statement through a read-only bank API and matches amount, reference and currency; an atomic settle-once step; three independent reviews before launch.

ResultPayments confirm automatically; the money risks the reviews raised were answered in follow-up releases; pre-launch review: all 46 files, 240 tests pass.

CS-13Next.js · Cloudflare Workers · D1 · payment QR · bank REST API · Vitest

Servers, updates and reliability

Nightly Shopware updates with a scripted checkout test on staging, then production

EU direct-to-consumer brand · Shopware 6 in Docker · staging twin

DeliveredAfter the nightly backup: an image rebuild when packages change, then scripted acceptance on staging and production — key pages, admin, cart add with readback, checkout confirm without payment, desktop and mobile screenshots. A weekly production-to-staging sync with mail delivery forced off.

ResultBoth environments “OVERALL PASS” on 2 October 2026, including cart readback and checkout confirm; package updates recorded night by night.

CS-03Shopware 6 · PHP 8.4-FPM · Docker BuildKit · Store API · Python · headless Chromium · Borg

Nightly server updates, one component at a time, each with a pre-verified rollback

Production Linux host · shops, help desk, marketing automation and AI services in Docker

DeliveredAn agent-run nightly lane that finds every installed component and records the previous version, a rollback command and a hash-checked copy before each change; databases and other stateful services move to a Sunday window, staging first.

ResultOne failure does not end the night: a failed item is rolled back along its pre-verified path and ticketed, and the run continues with the next item within its stop rules.

CS-29Debian · systemd · Docker Compose · SHA-256 rollback copies · database dumps · Borg

Backups that really restore: a 621-path coverage gap found and closed

Group of e-commerce companies on one dedicated server

DeliveredA recovery set twice a day, local and off-site Borg, and a restore drill that restores every dump into throwaway databases and checks the contents; coverage checked by real path; upgrades gated on the drill.

ResultFound and closed: 621 uncommitted paths whose only copy was the disk, and a hook that republished an incomplete recovery set while logging success.

CS-30BorgBackup · borgmatic · PostgreSQL dumps · systemd timers · Bash

Help-desk attachments back online after a container storage split

E-commerce support team · self-hosted help desk in Docker

DeliveredWorker paused; both stores and the database captured and merged into one shared mount that refuses to start without its source; lost files restored from other copies; rendering proved in a browser.

ResultAttachment check 404 → 200; 0 merge conflicts; 78 missing files restored except one; zero customer contact during the change.

CS-31Chatwoot · Rails · Sidekiq · Docker Compose · PostgreSQL · headless Chrome

AI agents and automation

Agents that do real work inside a business — customer support, refunds, returns, reporting — with guardrails, audit trails and tests. And LLM features whose output is structured, validated and measurable.

AI support desk on five channels: median first response 36 seconds

EU direct-to-consumer brand · e-mail, web chat, social and messaging apps · customers’ own languages

ProblemAnswers, returns, replacements and refunds every day, without an in-house support team.

DeliveredHelp-desk intake with one work item per conversation; a frontline and a lead agent working from written procedures; a reconcile job; a reply reservation against double replies; outgoing mail with signature and AI disclosure added by the send tool.

ResultMedian first response 36 seconds; 86 cases opened in 14 days; a platform outage of about 27 minutes with no customer reply lost; SPF, DKIM and DMARC pass.

CS-20Chatwoot · signed webhooks · agent gateway · LLM agents with fallback · Shopware Admin API

Refunds that AI agents can run safely

EU direct-to-consumer brand · cards and cash on delivery · several currencies

ProblemAgents had to refund returns and cancellations without over-refunding, paying twice or paying the wrong account.

DeliveredA guarded refund operation: exact order binding, the amount capped at the live refundable balance, one transaction per order, IBAN check per country, unknown outcomes sent to reconcile instead of retried.

ResultThe frontline agent is authorised to run it within written procedure; a real cancellation closed with the readback “order cancelled, transaction refunded, no tracking”.

CS-21Stripe API · Wise API · Shopware Admin API · Google Sheets · idempotency keys

Returns pipeline: photo intake, five carrier trackers, a twice-daily resolution cycle

EU direct-to-consumer brand · own warehouse · returns from three countries

ProblemReturns arrived through several carriers, warehouse photos and messages, logged by hand in a spreadsheet.

DeliveredScheduled parsers with dedupe, tracking reads for five carriers and a resolver that binds customer, order, carrier and physical evidence to the payment ledger, with caps and no-replay rules.

ResultFive carrier trackers live since 29 September 2026; every cycle confirms both backup receipts before it processes anything.

CS-22carrier tracking APIs · Google Sheets · Stripe · Wise · Shopware · scheduled agents

One tool gateway for AI agents: 1 432 operations across ads, analytics, shop, banking and carriers

Our own agent platform · a group of e-commerce companies

ProblemAgents received silently truncated reports, and credentials had to stay strictly per company.

DeliveredA Python service publishing YAML-declared operations to the agents, one company per deployment, and an audit of every report tool for caps, paging, totals and date windows.

Result1 432 operations from 147 declaration files, 1 100 of them without custom code; truncation fixed in 9 families with no change to any write behaviour.

CS-24Python · YAML · MCP · Google Ads · GA4 · Merchant Center · Microsoft Ads · Meta · Stripe · Wise

AI app with schema-validated model output and an evaluation harness

Consumer AI app · three languages

ProblemAI answers had to be structured, localised, safe in tone and testable — starting from an open-source prototype.

DeliveredA provider-neutral engine with versioned prompts, strict JSON Schema output validated on the server, a mock adapter, and an evaluation harness for locale, coverage, prompt leakage, length, latency and tokens.

ResultFirst commit on 24 September, release 0.1.41 on 3 October 2026; 288 tests passing, 0 type errors.

CS-25Next.js · TypeScript · Cloudflare Workers · D1 · OpenAI structured outputs · Vitest

Semantic search for an agent knowledge base, model chosen on real data

Our own agent platform · knowledge for five companies’ agents · Czech and English

ProblemThe search model had to be chosen on real company data, without per-token costs or lock-in.

DeliveredA frozen evaluation set from live data (679 duplicate pairs, 200 negatives), hosted and local models compared under a decision rule fixed before scoring, a held-out confirmation, then a lossless migration.

ResultA local 300M model chosen: 41 of 60 found in the top five vs 44 for the large hosted model (p = 0.51), 55 ms p95 query time; migration diff empty.

CS-26Python · EmbeddingGemma · Qwen3-Embedding · BGE-M3 · BM25 hybrid · McNemar test

Data and analysis

Pipelines, warehouses and studies that end in a decision: what the business keeps, what a price change does, when customers come back — with every number traceable to its source.

Owner-profit engine: what the business keeps per order after ads, fees, shipping and VAT

EU direct-to-consumer brand · several currencies · four ad channels

DeliveredA rerunnable pipeline from bank, payment and ad APIs, with customers anonymised, into analytics marts; profit scenarios, contribution per order and a weekly owner summary.

ResultWeekly summary live since 28 September 2026; complete coverage with estimated gaps of at most 1 %; an outdated VAT configuration found.

CS-33PostgreSQL · dbt · Airbyte · Python · Stripe · Wise · ad platform APIs

Pricing decision settled with data: elasticity study and competitor snapshot

EU direct-to-consumer brand · 8 markets

DeliveredMore than three years of orders: difference-in-differences against a control product family, regressions, break-even elasticity and a test design; competitor prices across 8 markets.

ResultElasticity ≈ −1.8 (range −1.2 to −2.6) against a break-even of ≈ −1.36: the past price cut most likely paid for itself, and an increase was not supported. 724 competitor price rows.

CS-34SQL · Python · pandas · statsmodels · public price data

Repeat-purchase research without exposing customer data

EU direct-to-consumer brand · two core markets

DeliveredA scripted pipeline over more than three years of orders that outputs aggregates only, with an independent in-database SQL cross-check and a personal-data scan.

ResultSecond purchase within 12 months 11.9 %, within 24 months 15.3 % (age-adjusted); every count reproduced independently; nothing sent to customers.

CS-35SQL · Python · numbered extraction and render scripts · PII scan

A paid 302-endpoint API turned into a verified, query-ready warehouse

Data project on a paid third-party calculation API

DeliveredAn endpoint coverage matrix, a request planner with stable hashes and a dry run, an idempotent resumable runner, raw-to-canonical normalisers, a SQLite query store, a verifier and a question-to-report command.

ResultAll 302 endpoints classified; every planned queue ran with zero pending requests; unique request coverage 259/260, 258/258, 204/204; verifier: 0 issues, 0 warnings.

CS-36Python · REST harvesting · JSONL · SQLite · CLI

More from our production work

Start with your goal

Tell us what you want to achieve. We reply within minutes with the questions that matter, or straight away with a plan: milestones, dates and a price. Nothing is paid to us up front — you see it working first.

Oldrich Works · Oldrich Dvorak stands behind every result