Tech Stack
The tools and technologies I rely on to build production-grade AI systems, open-source projects, and full-stack applications.
AI & Machine Learning
RAG Systems
End-to-end retrieval-augmented generation pipelines — chunking strategy, embedding selection, vector retrieval, and LLM integration tuned to the target corpus.
LLM Integration
Production LLM wiring with a focus on cost / quality / latency tradeoffs. Prompt engineering, structured outputs, and failure-tolerant orchestration.
AI Agents
Autonomous agent architectures with tool use, memory, and multi-step reasoning. Built for reliability in real deployments, not just demos.
Vector Databases
Embedding storage and semantic search at scale. Index design and retrieval tuning for high-precision document retrieval.
LLM Infrastructure
Serving, monitoring, and scaling LLM-powered systems. Observability patterns, latency budgets, and graceful degradation.
Multimodal Models
Applied multimodal LLMs to medical imaging — retinal scan analysis for early glaucoma detection. Bachelor's thesis graded 10/10.
Languages
Python
Primary language for AI systems, data pipelines, and backend services. FastAPI, LangChain, HuggingFace, and the broader ML ecosystem.
TypeScript
Full-stack TypeScript for front-end interfaces, React components, and Node.js tooling. Type-safe, maintainable, and production-ready.
Rust
Systems-level work where performance and memory safety matter. CLI tools and performance-critical components.
Frameworks & Infrastructure
FastAPI
High-performance Python API layer for AI backends. Async-first, type-annotated, and easy to document and deploy.
React
Component-driven UI with a focus on performance and accessibility. This site is built on Docusaurus (React) with custom components.
Document Processing
Ingestion pipelines for PDFs, structured documents, and institutional corpora. Chunking, parsing, and metadata extraction at scale.
Approach
The tooling choices listed here are always driven by the same principle: fit to the problem, not fit to the hype. Embedding models are chosen for domain quality, not leaderboard rank. Frameworks are chosen for reliability and iteration speed, not novelty. Every system is built to be observable, failure-tolerant, and built to iterate.