- RAG
- LLM
- FASTAPI
- DOCKER
- +1
AI Insight Assistant
An assistant that combines RAG and a SQL agent: it retrieves from a document corpus and queries a database, served with FastAPI and Streamlit and packaged in Docker. Code at github.com/delcenjo/ai-insight-assistant.
An internal assistant for a made-up company (Helios): you ask in plain language and it decides on its own where the answer should come from — the employee handbook or the database. Some questions are about written policy, some are about numbers, and the user should not have to know which is which.
- SourcesRAG + SQL
- ServingFastAPI + Streamlit
- PackagingDocker
Routing before answering
The interesting piece is the router: classifying the question and sending it to the RAG pipeline (semantic retrieval over the handbook, returning the chunk and its score) or to the SQL agent (read-only queries over the employee database). Two different problems behind a single text box.
Answers with evidence
When it answers from the handbook, it returns the exact chunk and its similarity score, not just the generated sentence. If retrieval finds nothing decent, saying so beats improvising: trust is earned by showing where every answer comes from.
Served like a product
FastAPI for the API, Streamlit for the interface and Docker to package it all. It is not a notebook demo: it comes up with a compose file and you try it from the browser.
Next projects:
Credit Risk Platform
End-to-end MLOps platform for credit-default scoring: reproducible training with a versioned model registry, a FastAPI inference API with input validation, PSI drift monitoring and a Prometheus and Grafana observability stack, all containerised with CI. Code at github.com/delcenjo/credit-risk-platform.
Transformer from scratch
A GPT-style language model implemented from scratch in PyTorch: multi-head attention, causal masking and residual blocks written by hand, plus a byte-pair tokenizer and an ablation study. Code at github.com/delcenjo/transformer-from-scratch.
Built with
