I build multi-agent automations and full-stack SaaS. Every system has a hard cost cap, monitoring, and real tests. You can open the code and check it.
Three automations, two SaaS products, and one AI media pipeline — all real, all on GitHub.
Multi-tenant salon SaaS. A PostgreSQL exclusion constraint makes double-booking impossible, and 5 cost-controlled Claude agents run the back office. Bilingual DE/EN, 327 tests, live demo.
View case study →Trilingual (TR/EN/DE) household budgeting SaaS with a strict prod/demo boundary. PWA, live, and in daily use.
View case study →Webhook lead intake on self-hosted n8n: validation, duplicate prevention, Airtable CRM, event logging, and Telegram alerts, with a separate error-handler workflow.
View case study →A Make scenario that reads Airtable CRM and log data, aggregates event-type counts, and sends a concise daily operations report to Telegram.
View case study →Webhook intake, validated Airtable record, a Gmail draft (not auto-sent), a Google Calendar follow-up, and a Telegram alert: one auditable Zapier workflow.
View case study →Multi-agent AI media pipeline with a hard $2.00 cost cap per run (threading.Lock), resume-from-checkpoint, and private-by-default publishing.
View case study →The work is deliberate — the systems are real, running, and on GitHub.
$ env --boundaries prod ● isolated secrets sealed demo ● seeded no real data preview ○ ephemeral per-branch $ stack --self-hosted docker running cloudflare-tunnel up no open ports cloudflare-access enforced logging on structured + retained
Webhook intake, validation, duplicate checks, CRM sync, and multi-channel notifications across n8n, Make, and Zapier. Self-hosted where it matters.
Structured JSON output, function calling, confidence scoring, and an orchestrator. For example, five Claude agents handle intake, booking, follow-up, and content.
Next.js 15, React 19, tRPC, Drizzle ORM, PostgreSQL/Supabase, and MySQL. Multi-tenant, multilingual, PWA-capable.
Hard per-run cost caps, monitoring, and confidence thresholds. AI can run unattended without overspending or writing bad data.
Docker, Cloudflare Tunnel and Access, Playwright, logging, and resume-from-checkpoint. Built to be inspected, stopped, and recovered.
I am building deeper retrieval and evaluation with LangGraph, LangChain, pgvector, Langfuse, and FastAPI. Status: in progress.
What the work takes by hand, and what each system does it in. (est.)
My career started in operations, not in code. For roughly eight years I ran field operations for events and festivals across Turkey — supervising crews of up to 13 and owning materials, staffing and truck logistics end to end for festivals of up to 15,000 people a day.
Then three years of last-mile delivery operations under Amazon Logistics in Vienna. Eight years of keeping real systems running under pressure taught me which problems are actually worth solving — and that most of them are just broken, manual workflows.
In early 2026 I moved into AI deliberately. I taught myself automation and SaaS development and started building the kind of systems I used to run by hand — with hard cost caps, monitoring and tests.
Today I’ve shipped six projects across four public repositories: workflow automations in n8n, Make and Zapier, an AI media pipeline, and two SaaS products — all open on GitHub. I’m now moving into AI professionally while continuing to build and release my own products.
Strict prod/demo API boundaries, secrets kept out of git history, and self-hosting where it matters — Docker, Cloudflare Tunnel and Access. You can click the public demos with no risk to any real data.
Here are the four failures I design against. This is why I show code instead of claims.
Lead intake, follow-ups, and reports done by hand break as soon as volume goes up. Most teams automate too late.
An automation with no logs, no cost limit, and no recovery path breaks without warning. You find out when a customer does.
A model call in a notebook is not a product. The hard part is getting it to run unattended with guardrails. That is where most projects stop.
A portfolio full of buzzwords proves nothing. The only real proof is code you can read and demos you can click.
___ _ _ ___ _____ ___ __ __
/ __| | | / __|_ _| __| \/ |
\__ \ |_| \__ \ | | | _|| |\/| |
|___/\_, |___/ |_| |___|_| |_|
|__/