Manuel Branco Nardi
SOFTWARE ENGINEER — APPLIED AI & SYSTEMS
I build software across the stack: from REST APIs and job queues to embedded firmware and AI-integrated real-time systems. Two first-author publications and a range of projects.
Profile / Rev. 2026
Publications
// 01Research on applying vision-language models to real-time, timing-constrained systems.
Projects
// 02Grouped by area: software development, embedded systems, and applied AI/robotics.
Workflow Automation Engine
A visual, n8n-inspired workflow automation tool: a React drag-and-drop step builder on top of an Express API, with a Redis/BullMQ job queue and independent worker processes executing steps asynchronously (including AI steps via the Anthropic API). Workflows are JSON definitions with step-to-step variable interpolation, stored in SQLite. Built iteratively, exploring queue-based system architecture with Claude Code as a development partner.
View on GitHub →Soul-Bound
A 3D horror-investigation game built with a team of five. Designed and implemented enemy AI behavior — detection, pursuit, and attack logic — and the checkpoint save/load system preserving player progress across the manor's three floors.
View on GitHub →LLM Lightshow
Dual-ESP32S3 architecture performing real-time audio FFT preprocessing and on-device LLM inference. Trained a compact model mapping frequency features to LED states, driving synchronized effects through custom ESP-IDF firmware — a full embedded ML pipeline from dataset to deployment.
View on GitHub →Arduino Musical Lightshow
A MIDI-synchronized Christmas light show: an Arduino reads note data from an SD card, maps each note to a shift-register output driving a relay-controlled light. Extended with a second circuit — linked over Bluetooth — using an ultrasonic sensor to trigger the show automatically as people walk by.
View on GitHub →Time-Constrained VLM Inference
Designed a reactor-based pipeline integrating semantic caching with vision-language models for autonomous driving context analysis, using the Lingua Franca coordination language to reason explicitly about timing constraints on VLM inference.
See publications →Project Kestrel
Autonomous drone project using the SORT multi-object tracking algorithm. Built and led development of a custom Gazebo simulation environment to test real-time tracking behavior against multiple moving targets, integrated with subsystem teams for full drone deployment.
View on GitHub →