Brazil (UTC-3) · remote

Pedro
Fioravante

AI Engineer: LLM agents, document AI, vision & speech.

I build software that does real work: LLM agents that act on codebases and dev workflows, vision and speech systems from dataset to deployed API, and the full-stack products around them. I care about what happens after the demo: tests, deploys, costs and the boring edge cases.

Pedro Fioravante, pixel artP1 · AI ENGINEER
What I work on

From model to product, end to end.

LLM agents in production

Tool-calling agents that read code, triage work, write fixes and open pull requests, with humans reviewing the result.

Claude APIMCPLangGraph

Document AI & RAG

OCR plus LLMs with structured outputs, in production at 30,000+ documents a month and 97% accuracy. RAG over scanned PDFs.

OCRStructured outputsLlamaIndex

Computer vision

Training CNNs from scratch and fine-tuning detection, segmentation and classification models. Datasets, benchmarks, APIs.

PyTorchYOLOCNNs

Speech

Real-time and batch Whisper transcription, on local GPUs and hosted inference, with voice activity detection and diarization.

Whisperwhisper.cppVAD

Local & private AI

LLMs and SLMs on llama.cpp for data that can't leave the machine, PII anonymization and GPU memory budgeting.

llama.cppCUDA

Systems & product

Rust services, Docker, CI/CD on Azure DevOps; Next.js, React, Supabase/Postgres with row-level security, offline-first PWAs.

RustNext.jsSupabase
Experience

Shipping GenAI to production since 2025.

  1. Statum
    Ribeirão Preto, Brazil · remote
    May 2025 – present
    AI Engineer (R&D)
    Aug 2025 – present
    • Designed and built an OCR + LLM document-extraction pipeline with structured outputs: 30,000+ tax documents a month in production at 97% accuracy, the rest routed to manual review.
    • LLM agents with tool calling (Claude API, MCP, LangGraph) that triage requests, read codebases and propose fixes as pull requests for human review.
    • Agents that act in real time during meetings: Whisper transcription, speaker diarization, low-latency responses.
    • Train and benchmark computer-vision models (CNNs, YOLO family) and serve them behind APIs.
    • Hosted and local models, including LLMs and SLMs on llama.cpp for sensitive data and PII anonymization; evaluation with LangSmith tracing and in-house test sets.
    • Ship Python (FastAPI), TypeScript (React, Next.js, Node.js) and Rust services with Docker and CI/CD on Azure DevOps.
    AI & Software Developer Intern
    May 2025
    • Natural-language-to-SQL agent with LangChain over MySQL, behind a containerized FastAPI backend.
    • RAG pipeline with LlamaIndex over multiple documents, including scanned PDFs through Tesseract OCR.
    • Compared LLMs for reliability and latency, including Llama 3 via Groq.
  2. Minas Gerais State University (UEMG)
    Brazil
    Sep 2023 – Oct 2024
    Undergraduate Researcher
    • Android farm-management app (Kotlin, Jetpack Compose) with a Node.js/Express backend and JWT auth.
  3. Language schools & private students
    Brazil
    2015 – 2023
    English Teacher
    • Eight years teaching English, which is why I'm comfortable working fully in English.

B.S. in Information Systems · Minas Gerais State University (UEMG), 2025 · Portuguese (native), English (fluent)

Projects

Things I've shipped in public.

Built for clients

Production systems, private code.

Freelance work alongside my main job. Happy to walk through either one.

Clinic scheduling SaaSprivate · 2026

Multi-tenant scheduling for dental clinics: drag-and-drop weekly calendar, patient records with audit trail, per-role and per-dentist permissions, files on Cloudflare R2, tenant isolation with Postgres row-level security. Installable PWA.

Next.jsSupabasePostgres RLSR2TDD
Attendance PWAprivate · 2026

Offline-first attendance for a factory floor: marks attendance with no connection, syncs through a queue, and enforces weekly rest-day rules.

Next.jsIndexedDBSupabaseTDD
Stack

Tools I reach for.

Languages
PythonTypeScriptRust
LLMs & agents
Claude APIAzure OpenAIGroqLangChainLangGraphLlamaIndexLangSmithMCP
Document AI
TesseractVLM OCRSentenceTransformersPinecone
ML
PyTorchHF Transformersscikit-learnYOLOWhisperllama.cpp
Product & infra
FastAPINext.jsReactNode.jsPostgreSQLSupabaseDockerAzure DevOpsLinux
Live

This page is served from a phone.

A Samsung Galaxy A13 on my desk runs a single Rust binary (axum + tokio). The numbers below are read from /proc and /sys once a second and streamed to you over Server-Sent Events.

connecting…
CPU
–%
CPU temp
–°C
RAM used
–
Battery
–%
Requests/s
–
Requests served
–
Server memory
–MB
Phone uptime
–
– watching now · server up –
Terminal

Prefer a shell?

Type help. Or skip the browser entirely: curl -s fioravante.dev.br prints a card with live numbers from the phone.

pedro@a13: ~
pedro@a13:~$
Contact

Have something that should actually work?

Agents, document AI, vision, speech or the product around them. Happy to walk through any of it.