AI & LLM Engineering
RAG pipelines, self-hosted models, vector search, and the infrastructure to run them — designed by someone building with the same stack.
From provisioning the server to shipping the RAG API — the full stack, handled with precision.
VPS provisioning, Docker orchestration, Coolify deployments, DNS management, SSL, firewall hardening. The server work most teams skip until it breaks.
Ollama on bare metal, model selection by hardware constraints, Qdrant vector stores, embedding pipelines. Private inference that doesn't bill per token.
Production-grade retrieval-augmented generation: hybrid search (dense + sparse), multi-provider LLM connectors, streaming, audit trails, and evals.
Systems built to run, not to demo. Each ships with Docker, CI, tests, and real auth.
Upload financial PDFs and query them in natural language. Hybrid search merges Qdrant vector results with PostgreSQL full-text via Reciprocal Rank Fusion. Multi-provider LLM connector supports Ollama, OpenAI, Groq, DeepSeek, Gemini and more. Audit trail with faithfulness and relevance scoring via Celery.
Ask any question across an ingested book library. APA-cited answers with exact page references. Handles PDFs without real page numbers via font-size heading detection. Incremental ingestion — only reprocesses changed books. Grafana + Loki observability stack included.
Intelligent document assistant for enterprise workflows. Conversational interface over structured and unstructured document sets. Built with multi-provider LLM support, streaming responses, and vector-powered retrieval.
Chosen for production reliability, not trend-chasing.
I build production LLM systems — from the server they run on to the API that serves them. No shortcuts, no toy demos.
Currently pursuing a Masters in Business Information Systems, with a research focus on applied LLM systems in enterprise contexts. The academic side informs how I think about system design; the production side keeps it grounded.
The niche is deliberate: AI infrastructure is where most LLM projects fail. Models are the easy part. Embeddings, retrieval quality, hybrid search tuning, latency at scale — that's where the work is.
RAG system, self-hosted AI setup, or LLM infrastructure consultation.