MSc Electrical and Computer Engineering
2023 – 2025 (completed)Instituto Superior Técnico, Universidade de Lisboa
Control, Robotics and AI | Computing Systems
Luís Marnoto
Instituto Superior Técnico, Universidade de Lisboa
Control, Robotics and AI | Computing Systems
Instituto Superior Técnico, Universidade de Lisboa
Generalized Referring Expression Segmentation on Aerial Photos
Status: Concluded
Submitted to IEEE J-STARS
Project page: luispl77.github.io/aerial-d | Paper
Honorable Mention by Accenture — Competition details
I am a technologist—I develop technology and believe that technology is the primary way humanity can grow. By developing technology to solve problems and improve people's lives, we advance as a species. I believe the future will be filled with machines, and I also believe in artificial intelligence. My long-term goal is to develop and help create the breakthroughs needed to achieve AGI.
I am working on a project called ContinuousML. It is my work toward that goal: building systems that can learn continuously during inference, build new understanding on the fly, and sustain long-horizon reasoning. This includes both new model/architecture work and evaluation efforts that test whether models can continuously learn during inference as they encounter new information.
I am also deeply passionate about education, which is why I am working on a project called EMWaver: an educational platform precisely for technology, covering software, firmware, and hardware—all the essential targets for understanding how machines work. As we move toward AGI and witness an explosion of technology, it becomes increasingly critical to understand technology at all levels—from high-level software abstractions down to the lowest-level electronics.
ContinuousML addresses one of the most fundamental limitations in modern AI systems: the inability to learn continuously during inference.
Current AI models, despite their impressive capabilities, lack the ability to learn new tasks or adapt to new information after their initial training phase. Once deployed, these models remain static—they cannot incorporate new knowledge or improve from experience. This is a critical gap that prevents AI systems from truly replicating human-like learning capabilities.
ContinuousML aims to solve this problem by developing models that can learn and adapt in real-time, continuously improving their performance and expanding their knowledge base during inference. This represents a paradigm shift from static, pre-trained models to dynamic, continuously evolving intelligent systems. Achieving Artificial General Intelligence (AGI) is another fundamental goal of ContinuousML, pushing toward machines that can understand, learn, and reason at a human level.
ContinuousML is currently closed source, and I am working toward creating a company focused on delivering breakthroughs in continuous learning.
As part of ContinuousML, I am developing KVGPT: an architecture aimed at achieving long-context understanding far beyond what current models can reliably do, while also making models substantially more efficient. The goal is for smaller models to approach the performance of much larger ones by using context more effectively.
This efficiency angle matters because I believe true continuous learning will require a personal model per user. That is not feasible with today’s trillion-parameter scale models, so it is imperative that we bring state-of-the-art language models to consumer hardware.
I am also working on Monte Lua, a benchmark within ContinuousML designed to evaluate how well models can build new “world models” at inference time—constructing fresh understanding and sustaining long-horizon reasoning as they encounter new information.
Monte Lua and KVGPT go hand-in-hand: the benchmark measures the kind of continual, long-range understanding that the architecture is meant to enable.
EMWaver is a platform consisting of a compact device called EMWaver, and a companion application (also called EMWaver) that runs on iOS, Android, and desktop. It is built to bring hardware learning and exploration to the masses—positioned as a modern replacement for both Arduino-style development and portable tools like Flipper Zero.
EMWaver is plug-and-play and intentionally fun: it comes prebuilt with features to explore and analyze the devices around you (only on hardware you own or have explicit permission to test).
Out of the box, it supports Sub-GHz signal recording, replay, cloning, and analysis directly in the app—so you can inspect protocols and reproduce captured signals. It also includes an infrared receiver and IR LEDs to interact with TVs and other infrared devices.
Beyond the built-in radios, EMWaver exposes GPIOs so you can connect external modules, sensors, and chips over standard interfaces like SPI, I2C, and UART—for example, nRF24L01+ (2.4 GHz) and RC522 (RFID/NFC), and many more.
The EMWaver app provides the built-in tools to interact with the device. On top of that, Wavelets are special in-app programs for the EMWaver platform that let you create custom functionality and workflows. Wavelets are designed for two things: (1) program and interact with EMWaver with zero building/flashing and no firmware workflow headaches—things “just work”; and (2) generate rich cross-platform user interfaces (buttons, text fields, pickers, and more) with consistent behavior across iOS, Android, and desktop.
This makes hardware interaction and “hacking” lightning fast: no compiling, no strange microcontroller error logs, and no time wasted wiring up ad-hoc UIs. If you can connect it to the GPIOs, you can extend EMWaver’s capabilities quickly.
The EMWaver board is 3.6 cm × 4.5 cm, compact and inexpensive, and it attaches directly to a phone via USB‑C for truly portable use—one of the key reasons it can serve as a practical alternative to Flipper Zero. By leveraging the client’s resources (CPU, memory, and UI), the system enables instant workflows and rich functionality for connected hardware.