Hello! I'm Pedro Landolt

Software Engineer — Security & AI. MSc in Software Engineering from the Faculty of Engineering, University of Porto, with a thesis (18/20) on adversarial attacks against LLM-based code review. I work at the intersection of application security and AI — from LLM red-teaming and adversarial ML to vulnerability management and backend platforms. I like understanding how systems fail and making them harder to break.

LinkedIn · Email · Hugging Face

Backend for an even/odd dice game, built as a technical exercise for an iGaming backend role. Players connect over WebSocket (with an HTTP API for automated testing), and the balance lives in a separate wallet, as in the seamless-wallet model used between casinos and game providers. Every bet is idempotent end to end, a wallet timeout triggers a rollback so a result is never shown without the money settled, and the protections are enforced by the database too: 20 concurrent bets for the same player let exactly one through. Ships with Docker Compose, a Postman collection that runs against the full stack in CI, and a partitioned append-only ledger.

Go · WebSockets · PostgreSQL · Docker · CI

Full-stack platform built end to end: ETL pipelines aggregating threat intelligence from CISA KEV, NVD, EPSS and VulnCheck, asset–vulnerability correlation, risk scoring, LLM-based remediation, and REST integrations with Jira and GitHub.

Python · ETL · PostgreSQL · LLM · REST APIs

Automated invoice processing for a logistics company: PDFs are extracted, validated against ERP vendor data, stored, and submitted to the ERP API — a full pipeline behind a role-based finance dashboard. FastAPI backend covering extraction, validation, ERP integration and JWT authentication, with a Vite/React frontend.

Python · FastAPI · React · ETL · PDF extraction

Operator dashboard for supervising a greenhouse spraying robot. The robot streams telemetry over WebSocket, and the UI treats connection state as a first-class safety concern, so the operator always knows whether the data is live. Includes a mission state machine, real-time push, a robot simulator, and 138 tests across a TypeScript monorepo.

TypeScript · React · WebSockets · real-time · testing

Agentic workflow built with LangChain/LangGraph, developed as part of the Devoteam challenge.

Python · LangGraph · LLM agents

Judge Evaluation via Selective Tactic-based Exploit Reinforcement. Black-box attack framework developed as part of my MSc thesis (18/20) that drives an LLM code-review judge to approve code that is provably wrong — an attempt only counts when the deterministic tests fail and the judge returns PASS.

Built on multi-armed bandits (UCB1, Thompson Sampling, KL-UCB, EXP3), it reached 67.6% first-attempt attack success — 2× the random baseline.

Dataset published on Hugging Face, with the full dissertation PDF included in the repository.

Python · LLMs · adversarial ML · reinforcement learning