Applied AI
AI integration, RAG & local models
Designing and building AI systems as a product: RAG pipelines, local models with Ollama, and ChromaDB as the vector store.
This area is different from AI as a development tool: there, AI helps me write code; here, AI is what gets built.
RAG pipeline at LIT
At LIT I design and build a set of AI microservices in Python/FastAPI that implement a full RAG (Retrieval-Augmented Generation) pipeline: ingestion, vectorization and query.
Local models, by design
I run the models with Ollama locally and use ChromaDB as the vector store. That’s not a technical limitation — it’s a decision: never send client data to external third-party APIs.
Infrastructure evaluation
Designing these systems takes infrastructure judgment, not just code: GPU architecture criteria, comparing MoE versus dense models, and calculating memory needs based on context size. This judgment is applied as part of the product design at LIT, without publishing the specific hardware used.
Product-driven design
The design always starts from real business needs — what documentation needs querying, how often, under what privacy requirements — before deciding on the technical architecture.