António Cruz

noted: collaborative MLOps platform with on-premises GenAILive

Self-hosted MLOps platform integrating MLflow, Airflow, DVC, Hydra, MinIO, ChromaDB and ArcadeDB into a single collaborative web interface. Real-time multi-user notebook editing, multi-runtime Python 3.10–3.14 (including free-threaded variants), kernel management, virtual environments, model registry and serving.

Native LLM tool-calling via 24+ MCP-style tool schemas (read and write tiers with approval flow for write actions); trigger-based skill registry auto-injects curated instructions into the prompt; per-Domain capability scoping (each Knowledge Domain pins its own tools and skills). GraphRAG (entity extraction, community detection, sameAs identity merging, denormalized chunk grounding) and Vector RAG (bge-m3 + cross-encoder reranking) over per-Domain knowledge bases.

Inference runs locally on Gemma 4 E4B via llama.cpp; the LLM router also supports Anthropic Claude as an optional backend.

FastAPI · Socket.IO · Python 3.10–3.14 · CUDA · Docker Compose · MLflow · Airflow · DVC · Hydra · ChromaDB · ArcadeDB · llama.cpp · bge-m3.

Demo video

Agent stack: privacy-first local AI orchestration

Three coordinated services that form a fully on-premises AI agent platform:

• agent_server: LLM control plane with an OpenAI-compatible /v1/chat/completions streaming facade. Local models behave as drop-in replacements for cloud APIs, hot-swappable per request via named presets (chat, judge, graph-answer, etc.).
• STT: Whisper Large v3 Turbo + Silero Voice Activity Detection over Socket.IO; real-time streaming transcription with multi-client support.
• TTS: Kokoro multi-language (English, Japanese, Mandarin, Spanish, French, Hindi, Italian, Portuguese) with WebSocket audio streaming, advanced sentence splitting via spaCy.

All inference on-prem. No external dependencies.

Python · FastAPI · Socket.IO · Hugging Face Transformers · llama.cpp · Whisper · Kokoro.

Autonomous warehouse robotics: voice + vision + LLM, end to end

End-to-end voice-controlled YouBot in Webots. The pipeline is: voice input, STT, quantized LLM (SmolLM2-135M via llama.cpp), structured JSON command, robot action, ResNet-18 box-damage classification, and TTS confirmation. Three-mode state machine (standby, patrol, goto) with aisle-aware path planning, dual-camera inspection, real-time inventory updates broadcast to connected clients.

Four independent Docker services communicating over Socket.IO. Entirely on-prem, no cloud calls in the loop.

PyTorch · ResNet-18 · Webots · llama.cpp · FastAPI · Socket.IO.

Demo: voice control Demo: autonomous patrol

scipredictor: multi-label classification across 148 research domainsLive

Multi-label classifier for academic papers across 148 research domains. Independent per-class probability output (no softmax constraint) so an interdisciplinary paper can simultaneously belong to multiple fields, mirroring real research practice instead of forcing a single label. Companion case study using SciBERT for arXiv primary-subject prediction with hierarchical-domain analysis and confidence-based triage of ambiguous cases for human review.

PyTorch · Hugging Face Transformers · SciBERT.

GAN comparative study + GAN-vs-Human gameLive

Conditional MNIST generators benchmarked across four loss strategies (BCE, LSGAN, Hinge, WGAN-GP) with identical architectures and FID and KID evaluation. The best-performing model powers a real-time browser game scoring human-drawn vs generated digits via a trained MNIST classifier. Three difficulty levels with time pressure (2.0 s, 1.0 s, 0.5 s).

PyTorch · WGAN-GP · FID and KID · WebSockets.

Multi-step temperature forecasting: GRU vs Transformer

Multivariate-input, univariate-output forecasting study on the Jena Climate dataset (10-minute resolution, 2009–2016). Predicts 24-hour air temperature using GRU and Transformer architectures under controlled conditions: temporal resampling, multi-step windowing, multi-horizon evaluation, hyperparameter exploration. Six input features (temperature, pressure, humidity, wind speed, max wind speed, wind direction), each paired with timestamps.

PyTorch · Keras · time-series preprocessing.

Report

DQN vs PPO on LunarLander-v3: value-based vs policy-based RL

Comparative deep-RL study contrasting Deep Q-Network and Proximal Policy Optimization in the LunarLander-v3 environment. Identical environment settings and multiple random seeds; analysis combines learning curves, sample efficiency, cross-seed stability and qualitative landing-behavior evaluation from recorded simulations. Covers DQN's overestimation bias, ε-greedy exploration, replay buffer dynamics; PPO's clipping, GAE, and entropy-driven exploration.

PyTorch · Gymnasium · DQN · PPO · GAE.

Report + annexes

FEM execution-strategy benchmark: CPU vs GPU vs JITLive

Same Finite Element Method problem implemented across multiple execution backends (sequential CPU, multi-process and shared-memory CPU parallelism, Numba JIT, Numba CUDA, CuPy with custom raw kernels) under identical numerical formulation, discretisation, boundary conditions and solver. Numerical equivalence preserved across all paths, enabling apples-to-apples comparison of execution behaviour, performance and scalability. Containerised with platform-specific helpers and automatic CPU and GPU fallback.

Python · NumPy · Numba · CUDA · CuPy · Docker.

doco + docbro: document toolingdocodocbro

• doco: Self-hosted notebook to styled-Word converter, 5-step wizard, real-time WebSocket progress. Configurable fonts, layout, tables, code styling, image handling, optional HTML and Markdown exports.
• docbro: Static client-side documentation browser. Split-pane tree navigation, tabbed switching, KaTeX math, Mermaid diagrams, syntax highlighting, image lightbox, deep linking, scroll-synced TOC.

Python · FastAPI · WebSockets · pandoc · KaTeX · Mermaid.

Chest X-Ray pneumonia classifier: binary medical-image CNNLive

Convolutional neural network for binary classification of chest X-ray images (pneumonia vs. normal). Full training pipeline with data augmentation, validation, and evaluation by accuracy, precision, recall and F1.

Python · PyTorch · CNN · medical imaging.

OpenCV Lab: interactive image-processing sandboxLive

Web-based experimentation environment for classical computer-vision techniques: grayscale and colour-space conversions, resizing, rotation, Gaussian and median blurring, Canny and Sobel edge detection, binary and Otsu thresholding. High-performance aiohttp backend with non-blocking I/O; Socket.IO for real-time browser interaction.

Python · OpenCV · aiohttp · Socket.IO · WebSockets.

Home-lab production stack: the platform behind the portfolioLive

Personal infrastructure hosting the entire portfolio. ~15+ Docker containers across noted's MLOps stack, the agent stack (STT, TTS, orchestrator), the application gallery and the web frontend. Linux on WSL2, nginx TLS facade with Let's Encrypt, dynamic-DNS public address. Services auto-start at host boot, the site is available even when the developer is signed out.

Fully on-premises AI stack (Whisper, Kokoro, Gemma, bge-m3, bge-reranker, ArcadeDB, ChromaDB), no cloud-AI API dependency.

WSL2 · Docker Compose · nginx · Let's Encrypt · NoIP · systemd.