Selected projects from recent years, three different industries, one principle: I build automation as infrastructure, not marketing slides. Numbers shown are illustrative; they convey scale and impact, not exact operational statistics.
Real projects & websites
Own tool · free
Receipt Maker
Receipt and invoice generator running entirely in the browser, nothing is uploaded. Thermal 58 and 80 mm rolls, A4, per line VAT with a summary, Code 128 and EAN-13 barcodes, payment QR codes. Hand written QR and Reed-Solomon encoders, zero dependencies, 196 tests.
Website for a barbershop in Sydney (East Hills, NSW). Premium dark look, services and pricing, plus online booking through Cal.com, walk in or book in a couple of taps.
Concept website for a family butcher in Padstow (NSW, Sydney). Order your cuts from your phone, say how you want them trimmed, pick them up at the counter without queueing. Built on spec as a demonstration, the shop is not a client.
A side project in development. Details coming soon.
In development
Launching soon →
Sample case studies
Illustrative model, not a delivered engagementLogisticsData pipelinesScope: ~6 weeks
Invoicing automation
A global logistics company processed international invoicing by hand: a team of four re-typed data between the TMS, the customs database and accounting. Every shipment waited on paperwork, 48 hours on average.
Starting point
4 full-time employees just re-typing data between three systems
~3% manual re-typing error rate, every error meant a customs delay
Invoicing capped growth: more shipments = more people
Solution
I built event-driven orchestration on custom middleware: every event in the TMS (loading, border crossing, delivery) automatically triggers data validation, customs document generation and posting to accounting. It connects to the existing systems via API, so the company didn't have to change software. People handle only the exceptions the pipeline flags for manual review.
Connected to the existing TMS and accounting system via API, no software replacement
A validation layer with rules for 14 countries
An exceptions dashboard, the team sees only cases that need a decision
The figures above are modelled, not measured at a client. They show the scale and type of impact, not the operating statistics of a delivered engagement. Documented figures appear only in the Galochrom study above.
Illustrative model, not a delivered engagementBankingAI agentScope: ~8 weeks
Internal AI assistant
A bank's internal helpdesk received over 900 queries a day, from password resets to interpreting internal policies. Twelve operators dug through thousands of pages of documentation for answers. Public AI tools were unusable due to regulation.
Starting point
Average response time of 4.5 hours, with daytime peaks
Strict requirements: no data may leave the bank's infrastructure
Knowledge scattered across directives, wikis and e-mail threads
Solution
I deployed a stateful conversational agent on RAG architecture, running entirely in the bank's private cloud. It was trained exclusively on internal documentation; it backs answers with source citations and hands complex cases to humans with the full conversation context.
A private vector database, APIs with no data retention
Every answer linked to its source directive, auditability
Escalation protocol: uncertainty → handover to an operator, never a guess
Solution preview
internal-assistant, private cloud● online
What's the limit for instant payments on business accounts?
The instant payment limit for business accounts is CZK 2,500,000 per transaction. Payments above this limit automatically fall back to standard processing.Source: Directive PL-2024/07 · art. 4 par. 2
And who can raise the limit for a client?
Limit increases are approved by the risk department at the request of the client's banker. The request form is in the system under RM-07. Shall I hand this over to an operator?Source: Directive RM-2023/11 · art. 7
⌀ response: 40 s78% without escalation● data never leaves the bank
The figures above are modelled, not measured at a client. They show the scale and type of impact, not the operating statistics of a delivered engagement. Documented figures appear only in the Galochrom study above.
Illustrative model, not a delivered engagementSaaSPredictive middlewareScope: ~5 weeks
Customer churn prediction
A fleet-management SaaS platform was losing customers quietly: when a client stopped using the product, the team found out only from the cancellation notice. Customer success reacted late and indiscriminately.
Starting point
Monthly churn of 4.1%, growing with the customer base
Churn signals sat in telemetry nobody evaluated
Customer success called customers blindly
Solution
I designed predictive middleware that reads product telemetry in real time, scores each account's churn risk and triggers graduated interventions on its own: from a personalized onboarding e-mail to a task for the account manager with a precise description of what's happening with the client.
Connected to product telemetry, CRM and e-mailing via API
A risk model retrained weekly on fresh data
Interventions escalate by score, automation handles light cases, people the critical ones
Solution preview
churn-radar, account risk scores● live telemetry
AccountChurn riskUsage trendAutomatic action
Fleetio Trans18stable, no action
Logimax s.r.o.54→ onboarding e-mail sent
TransCargo CZ87→ task for the account manager
Detection lead time: 0 daysMonitored: 0 accounts● model retrained 2 days ago
Churn risk dashboard preview (illustrative data).
Result
−31%customer churn per quarter
9 daysaverage risk detection lead time
5×higher retention campaign reach
The figures above are modelled, not measured at a client. They show the scale and type of impact, not the operating statistics of a delivered engagement. Documented figures appear only in the Galochrom study above.
Dealing with something similar? Write to me, I'll gladly look at your case and tell you whether and how it can be automated.