In progress
Ask Alfonso
A small language model built from scratch (BPE tokenizer, decoder-only transformer and retrieval) that answers questions about my work.
- My role
- Solo project
- Team
- Solo project
- Year
- 2026
The problem
Recruiters have specific questions. A chatbot can answer them, but only if it never makes things up.
What I built
- A byte-pair-encoding tokenizer written from scratch.
- A decoder-only transformer: embeddings, causal multi-head self-attention, layer norm, MLP and residuals.
- TF-IDF retrieval that grounds every answer in this site's data, with a fallback when confidence is low.