Skip to content
Edson CostaIntegration & Applied AI Engineer

AI system design

A recruiter-facing RAG system with evidence, not guesswork

This portfolio assistant combines hybrid retrieval, confidence gating, source allowlisting, multilingual evaluation, and privacy-aware streaming.

Retrieval and generation flow

  1. 01

    Verified career profile

  2. 02

    Gemini Embedding 2 · 768 dimensions + multilingual BM25

  3. 03

    Weighted RRF · semantic 0.6 · lexical 0.4 · k=60

  4. 04

    Confidence + confidentiality gate · initial threshold 0.55

  5. 05

    Gemini 3.6 Flash · streaming · store=false

  6. 06

    Allowlisted source cards + signed five-answer budget

eval.golden_set

Evaluation evidence

Live Gemini evaluation is pending a restricted API key. Targets are not presented as passed results.

  • 30 committed cases · 10 each in EN, PT-BR, and ES

  • Target: at least 90% expected-source recall in the top five, per language

  • Target: 100% rejection of unknown, confidential, and prompt-injection fixtures

  • Sentence-level grounded-answer checks against allowlisted evidence

Privacy by design

  • No transcript database; the conversation exists only in sessionStorage.
  • Analytics and operational logs never include questions, answers, prompts, transcripts, or source text.
  • Only verified public career sources can be retrieved and cited.
  • A signed server-side token enforces the five-answer limit without user identity.