João Pedro Maia Contact
Machine Learning Engineer · Software Engineer

João Pedro Maia

I build LLM agent systems and the infrastructure underneath them.

Based in Copenhagen, working remotely for aiXplain. MSc Data Science at the IT University of Copenhagen. Open to conversations about machine learning engineering, LLM agent systems and the infrastructure under them.

Experience

aiXplain

2024 – present
Machine Learning Engineer · Remote (Copenhagen)

Agent systems and the infrastructure under them: an agent that builds agents, LLM routing, human-in-the-loop approval and voice.

Key projects

Led the second version. Users describe what they need in a conversation. A classifier reads the whole thread and routes it to search, agent-building, debugging or chat agents, and asks a clarifying question when it can't tell. It runs on Kubernetes (FastAPI, Celery, Redis), with tools exposed over REST and MCP.

Python · FastAPI · Celery · Redis · Kubernetes · MCP

Cost-aware LLM gateway

2026

Designed and directed a Rust LLM gateway. It speaks the OpenAI and Anthropic APIs, signs cost receipts with Ed25519, and stores no prompts. A routing policy goes live only after a pre-registered non-inferiority test passes. It runs in production behind the company's coding agent.

Rust · Ed25519 · LLM evaluation

Human-in-the-loop approval system

2026

Wrote the RFCs for, and directed the implementation of, human-in-the-loop for agents. An agent can ask the user a question or wait for approval before a tool with side effects runs: payments, CRM writes, or emails. Paused runs park on Temporal and resume from Redis checkpoints after signed, one-time decisions.

Python · Temporal · Redis

Realtime voice agents

2026

Built a realtime audio SDK and designed and directed a cascaded voice-agent engine (voice-activity detection, speech recognition, the agent and text-to-speech), with barge-in so users can interrupt.

Python · Voice AI

A one-click Mac app for the platform

2026

Packaged the company's multi-container platform as a one-click local Mac app in three days, then maintained it. It needs no Docker Desktop and no terminal, and agents can reach mail, calendar, files and shell through a host gateway with opt-in write access.

Swift · Docker

A Swift SDK

2024 – 2025

Proposed and built the company's first SDK outside Python, in Swift for iOS, macOS, watchOS and tvOS, with no brief. It powers the company's iOS chat app and an open-source agent-builder app for macOS.

Swift · iOS · macOS

SolarCore

2025 – 2026
Founder

Founded a solar monitoring platform. Installers could see their customers' plant output and power bills in one view. Designed collectors that sign in to three inverter-vendor portals with no simple public API (Playwright, SQS, ECS Fargate). Automated utility-bill downloads, including CAPTCHAs and password-protected PDFs, with LLM parsing. Built a multi-agent assistant that answered questions on live data over WhatsApp, Slack and the web. Led a four-person development team from first prototype to a sale. The company is now closed.

Python · Playwright · PydanticAI · AWS (Lambda, ECS Fargate, RDS/PostGIS) · Terraform

PUC-Rio

2023 – 2025
Researcher

Final-paper research on mammography segmentation with unbalanced datasets (U-Net, SegFormer). It led to a co-authored journal paper in Signal, Image and Video Processing (2025).

Apple Developer Academy, PUC-Rio

2022 – 2023
iOS Developer (program participant)

Built iOS and watchOS apps in teams of up to 40 people, and published apps on the App Store. Projects included AI photo restoration with a model-serving REST API, an astronomy app for iOS and watchOS, and WeatherKit alerts with widgets. Helped with ML lectures and the competitive-programming study group.

Quantum Finance

2021
Software Engineer Intern

Remote internship at a fintech company (January–July 2021).

Education

MSc Data Science, IT University of Copenhagen

2026 – 2028 (expected)

Advanced Applied Statistics, Algorithm Design, Data in the Wild, Seminars in Data Science. Current seminar: reproducing a NAACL 2025 study on LLM prompt sensitivity on the university's HPC cluster.

BSc Computer Science, PUC-Rio

2019 – 2024

Department of Informatics, whose graduate program holds CAPES grade 7, the highest grade in Brazil's national evaluation. Final paper with the Tecgraf Institute on deep-learning segmentation of mammography images that still works on images from other scanner vendors, graded 10/10. It became a co-authored journal paper in Signal, Image and Video Processing (2025).

Publications

VeriEvolve: Evaluator-First Evolution for Tool-Using Agents

Y. Li, K. A. Yuksel, G. Tur, D. Hakkani-Tür, J. P. Maia

REALM Workshop at EMNLP 2026 · Budapest · Accepted · OpenReview

LLM-Driven Correlation-Aware Tournaments

K. Yuksel, J. P. Maia, T. Castro Ferreira, M. Al-Badrashiny

ICLR 2026 Workshop on Advances in Financial AI · OpenReview

LLM-Driven Active Listwise Tournaments for Portfolio Selection in Large Asset Universes

K. Yuksel, J. P. Maia, T. Castro Ferreira, M. Al-Badrashiny

ICLR 2026 Workshop on Advances in Financial AI · OpenReview

AlphaLoss: LLM-Driven Evolution of Robust and Interpretable Portfolio Optimization Objectives

K. Yuksel, J. P. Maia, T. Castro Ferreira, M. Al-Badrashiny

ICLR 2026 Workshop on Advances in Financial AI · OpenReview

Improving the generalization of deep learning models in the segmentation of mammography images

J. Hurtado, J. P. Maia, C. A. Sierra-Franco, A. Raposo

Signal, Image and Video Processing 19(18):1451 · Springer · 2025 · DOI · arXiv

Recognition

Attended WWDC24 at Apple Park

2024

WWDC23 Swift Student Challenge Winner

2023

Apple. Built Night Canvas, an AR playground that matches a hand-drawn constellation to the real constellations using spectral graph matching.

Writing

A step-by-step guide to running an AI agent on a rented server with zero-trust security: key-only SSH, Tailscale private networking, systemd sandboxing, and automatic patching.

All writing