Built for clients in Australia, Europe, and the US, and most of them are still running today, answering real customers at three in the morning. Voice agents, document systems, automations, and the messy integration work in between that decides whether any of it holds up.
One booking and dispatch scenario. ~2,000 inbound calls a month, six technicians, every failure path handled.
A place for lawyers and doctors to upload their own files and just ask questions about them. Every answer points back to the exact page it came from, so nobody has to take the AI's word for it. Runs on AWS with per-user access, a full audit log, and retention rules that keep both HIPAA and GDPR happy.
A German-language phone system that answers the usual questions first, then books the customer in with one of six technicians or the delivery team. It checks real availability, writes to the calendar, sends the confirmation text, and knows what to do when a slot vanishes halfway through the call.
A recruiting platform that rings candidates on its own, works out who is worth talking to, and passes those ones to a human. I took it from a blank page all the way to something people can sign up to and use.
vonrel.com
A booking site for a farm stay in northern Poland, built and shipped for the client in both Polish and English. Not everything is backend plumbing. When the job needs something customers actually look at, I build that too.
ranczonadstawem.com
A new lead comes in, gets researched, gets called and qualified by a voice agent, then moves through the CRM with the proposal, the follow-up emails, and the status updates all written as it goes.
Pulls files from S3, chunks and embeds them, and serves the most relevant passages before the model answers. Built, deployed, and tuned solo.
How I actually build now. Custom agents, skills, and commands wired into Claude Code so it knows the codebase and the rules of the business it is working on before I ask it anything. Same setup runs my accounting: it pulls transactions from Stripe, Wise, and Gmail, categorises them, reconciles the ledger, and writes the quarterly tax reports.
A coaching and progress-tracking app built full stack for my own students, giving them sprints, structure, and live feedback while they build their own AI systems.
vaib-pods.vercel.appAgents that take and make real phone calls for sales, support, and bookings. I set up the phone lines underneath them too, which is usually where these projects come unstuck.
Taking a job someone does by hand every day and turning it into something that runs itself. I plan for what happens when it breaks, not just when it works.
Systems that answer questions from a company's own documents and get it right, with the answer traced back to the page it came from.
The bit most AI projects live or die on. Getting reliable output every time, testing it properly, and keeping it working as models change underneath you.
I work in Claude Code and Codex every day. It is the reason a project that used to take months now takes a couple of days.
Keeping European client data in Europe, and building on AWS to a standard US healthcare clients can sign off on.
I taught high school maths before any of this, and the same skill turns out to decide whether a technical project lands. I run weekly live sessions where people build real systems alongside me, I have trained over 300 paying students, and I speak at events like the AfricAI Automation Mastermind in Cape Town. In practice it means I can sit with someone non-technical and have them actually understand the trade-off, rather than just nod at it.
A weekly live session. Everyone builds the thing on screen with me, rather than watching.