I build the parts of a product that users never see but always feel: the services that take messages from six different channels and route them in order, the queues that retry what failed and park what cannot be saved, and the pipelines that turn messy documents into something an LLM can actually answer from.
For about five years I have been working mostly on the backend side with Node.js (NestJS) and Python (FastAPI), on multi-tenant and event-driven systems that run in production, plus enough Vue.js to ship the screens on top of them. Lately most of my time goes into LLM and RAG work: parsing, chunking, embeddings, vector search and wiring several model providers behind one interface.
|
Tenant-isolated environments, tenant-scoped caching, configuration that can be moved between tenants and collections, REST APIs with versioning and OpenAPI docs. Authentication with JWT and refresh tokens, authorization with attribute-based rules (CASL). |
A unified messaging layer for WhatsApp, Telegram, MS Teams, Slack, Facebook and Discord. Webhook ingestion, asynchronous processing, intent / entity / FAQ management, LLM-powered flow steps, speech-to-text and text-to-speech, at loads of a few thousand conversations a day. |
|
Document parsing that reads structure (headings, font sizes), semantic, recursive and size-based chunking, embedding and enrichment steps (including LLM-based table enrichment), similarity search tuning in Qdrant, and a provider layer over OpenAI, Anthropic, Google, Hugging Face and others. |
RabbitMQ with retries, dead-letter, priority and delayed queues; in-process domain events in modular monoliths; scheduled jobs, backup / restore and real-time analytics. I like systems where every message has a known path, including the failure path. |
|
OCR on scanned PDFs (Tesseract, OCRmyPDF) with rule-based segmentation, so documents with many unrelated records on one page can be split, filtered and handed to an LLM for structured extraction. |
Unit, integration and end-to-end tests (Jest, Vitest, tests against a real PostgreSQL), raising coverage on legacy code to 80%, refactoring old codebases, and CI/CD with Jenkins and Argo CD onto Kubernetes. |
| Area | Tools |
|---|---|
| Languages | |
| Backend | |
| AI / LLM | |
| Data & messaging | |
| DevOps | |
| Testing | |
| Frontend | |
| Ways of working |
|
Does what I tested match what actually happens? A trading-strategy research platform: hourly market data, 11 indicators feeding rule-based strategies, an event-driven backtest engine with fees, slippage, walk-forward and Monte Carlo analysis, and paper portfolios (internal and Alpaca paper). The backtest and the paper portfolio share one execution engine, so a divergence report can prove the two
produce identical trades. Modular monolith with pure domain layers, ports and adapters, |
Can I check where the answer came from? Ask your PDF, Markdown and text files a question in English or Turkish and get an answer in the same language, where every claim carries a numbered citation that opens the exact file, page and passage. If the answer is not in the documents, it says so. Files are parsed, chunked by tokens and embedded into PostgreSQL + pgvector (one partial HNSW index per embedding size), grouped into collections with their own models. OpenAI and local Ollama sit behind one provider layer, a retrieval panel shows every passage found and whether it was used, and an eval set measures retrieval quality. Vue 3 interface, API keys stored encrypted. |
- One code path for the same decision. If two features answer the same question, they should run the same code, so their results can be compared instead of argued about.
- Plan for the failure path. Retries, dead-letter queues, idempotent ids, kill switches and audit logs are part of the design, not an afterthought.
- Keep the domain pure. Business rules without framework imports, outside services behind ports, so the core is easy to test and easy to change.
- Test against the real thing where it matters. End-to-end tests on a real database, fakes only at the network edge.
- Be honest about limits. Every project README lists what it does not do yet.