Gabriel Chaves

Browser extension that fills job application forms through an LLM — designed so the model cannot invent facts about you.

• Chrome MV3 and Firefox extension that maps arbitrary job application forms to a structured candidate profile through an LLM, with no per-site adapters: the form is scraped for labels, options and required flags, and the model resolves the mapping.
• Designed against model fabrication: the LLM returns a fill plan that is schema-validated with one corrective retry before any field is written, open-ended answers are anchored to the profile and the job posting, and no code path submits the form.
• Local-first, no backend: API key and profile stay in extension storage, CVs are parsed in-browser with pdf.js, and any OpenAI-compatible endpoint works, including a local Ollama.
• GitHub Actions CI with lint, type checks, a Vitest suite and both browser builds; a containerized six-page ATS test environment deployed to Cloud Run with Terraform.

• TypeScript
• Chrome MV3
• Firefox
• Vitest
• Terraform
• Cloud Run

Citation-grounded RAG over Brazilian tenancy law — every claim verified in code against the statute before it is shown.

• Answers questions against the full text of Lei 8.245/1991: each claim must be backed by a verbatim quote from the actual article, checked by normalized substring match in Python — not taken on the model's word — or the answer is flagged.
• Two Gemini calls by design: one streams a responsive answer over SSE, a second structured-output call extracts {article, quote} pairs for verification — avoiding the fragile alternative of parsing citation tags out of a live token stream.
• Ingestion scrapes planalto.gov.br and strips both strikethrough forms of revoked amendments (HTML tags and CSS line-through) so historical wording never ships as current law; the chunked output is committed as a reproducible seed.
• FastAPI + Vue 3 on a single Cloud Run service, Supabase pgvector as the store, Terraform for the GCP resources and keyless deploys via Workload Identity Federation.

• Python
• FastAPI
• Vue 3
• Gemini
• pgvector
• Terraform
• Cloud Run

Private RAG chat over an Obsidian vault, with hybrid retrieval and git-versioned write-back.

• Hybrid local retrieval — e5 embeddings plus BM25 — over the vault's notes, re-indexed automatically after every write.
• The assistant writes knowledge back into the vault as proper PARA notes with wikilinks and frontmatter, and every change is a git commit you can undo.
• Tool-use agent loop with on-demand SKILL.md skills, DeepSeek server-side web search with required source citations, and SSE streaming behind a FastAPI service.

• Python
• FastAPI
• DeepSeek
• e5 + BM25
• SSE
• Cloud Run