LLMs, local agents, and orchestration

psychic-waddle

Portfolio RAG assistant backend for recruiter-facing questions about my public profile.

Public codeProduction assistant

Overview

psychic-waddle is the retrieval-backed assistant that powers the chat overlay on this portfolio site. It searches reviewed knowledge about my public profile, applies answerability rules, and uses a configured LLM only to phrase supported answers in a professional style.

Role in this site

Visitors talk to the assistant through same-origin /api/assistant/* routes on pigreco.xyz. The static site and Nginx edge handle the browser contract; this repository owns retrieval, policy, and answer generation. The frontend does not embed CV facts or call the backend origin directly.

Deployment

The production backend runs on a dedicated 3-vCPU VPS with 8 GB RAM. That is enough for PostgreSQL, embedding search, and local model inference, but not infinite capacity. Responses can take noticeable time while retrieval and generation complete, especially under load or on longer questions. Please be patient when waiting for an answer.

Design principles

  • Reviewed knowledge first; the model does not invent profile facts.
  • Refusals and clarifications when context is missing, ambiguous, or off-topic.
  • No visitor tracking or treating anonymous chat logs as ground truth.