Doga O.

I build automation that replaces manual work — not fragile scripts that need babysitting, but production workflows that run unattended, handle failures loudly, and hand your team back hours every week. One example: I built a system that answers questions from a company's own documents and data, and refuses to answer when the source isn't there instead of inventing something. You ask in plain English, and every answer comes back with the source behind it. If your team keeps answering the same questions out of scattered docs and spreadsheets, that's exactly the work I do. Where I come from: Before engineering I spent years in finance and accounting at BASF and PSD Bank. I didn't learn about repetitive back-office work from a client brief — I did it myself. That's why my automations map to how the work actually flows, not how a diagram says it should. What I deliver: Multi-agent systems that route and self-correct (LangGraph): a supervisor reads each question and hands it to the right specialist — an exact database lookup, a recommendation, a document question, or an honest refusal when it's out of scope. Before answering, a grader checks whether what it retrieved actually addresses the question; if it doesn't, the system rewrites the search and goes round again. That retry loop is why it's a graph and not a chain. Conversations are stored per thread in Postgres, so follow-ups work across sessions. Public on GitHub. Answering questions from your own documents (RAG): a retrieval system that checks every quote is really in the source before showing it, so it doesn't invent things. It searches three ways at once — exact database filters, semantic vector search, and fuzzy text matching — so it handles both precise constraints and loose, natural questions. If the answer isn't in the sources, it says so instead of guessing. Public on GitHub. A Telegram chatbot over that same backend: it turns plain-English requests like "completed manga with 100+ chapters" into real database filters, shows the source behind every answer, and hands off to a human rather than guess when it can't help. A WhatsApp Cloud API version is built and tested. Getting systems into production and keeping them there: I took an internal ERP from a developer's machine to a live on-premise deployment — a Docker Compose stack with health-gated startup, a Traefik reverse proxy isolating the database from the edge, Active Directory login, and nightly backups. Two things from that job are why I work the way I do. I ran a full restore against a real backup rather than assuming it would work, and found a fault that would have left the system down during a recovery. And before switching on AD login, I checked what it would do, found it would have auto-created accounts for all 63 people in the directory, and restricted it first. Intelligent email triage: a Gmail pipeline that classifies every incoming message through Claude and routes it automatically. The team stops reading everything and only sees what needs a human decision. Research and monitoring pipelines: a system that runs continuously across six marketplaces, parses messy seller titles, matches them against a 20,000-item database, calls an OpenAI vision model when the text isn't enough, and scores every find so the best opportunities surface first. It runs on a schedule and reports on its own. Custom AI integrations (MCP): I built an MCP server that connects Claude directly to AutoCAD's internal COM API — you instruct Claude in plain language and it creates layouts, inserts blocks, applies text styles, and captures screenshots, executing inside AutoCAD in real time. The COM API is poorly documented and timing-sensitive, so getting it stable was the hard part; it's published openly on GitHub. Beyond AutoCAD, I design MCP integrations that let Claude operate real software safely and reliably — internal tools, desktop applications, and legacy APIs nobody has connected AI to before. Cross-platform mobile apps: production Flutter apps with real-time multi-device sync via Firebase Firestore and Riverpod, offline-first SQLite persistence, and live third-party API integration over REST and GraphQL. How I work: Failures surface immediately, never silently. If something breaks, you know within minutes — not three days later when data goes missing. Everything is documented and handed over so it stays maintainable: you own the system after I leave. Error handling is built in from day one, not patched in after the first incident. I use Claude Code as a core part of my build process. Stack: n8n · Claude API · OpenAI API · Claude Code · Anthropic MCP · LangGraph · pgvector · FastAPI · PostgreSQL · Docker · Docker Compose · Traefik · Linux · Bash · LDAP/Active Directory · Flask · Flutter · Dart · REST APIs · GraphQL · SQLite · Firebase/Firestore · Google Workspace APIs · Microsoft Graph API · Webhooks · OAuth · JavaScript Languages: English and German fluently, Turkish natively.