We build machine learning systems grounded in data and validated by research. From satellite imagery to archival documents, we work where the data is messy and the impact is tangible.
Deep learning for environmental prediction, document understanding, and conversational AI. From data collection to deployed inference.
Multi-source pipelines turning satellite imagery, climate reanalysis, and archival records into structured, model-ready features.
GPU clusters, Kubernetes orchestration, and MLOps pipelines. We build and operate the infrastructure our models need.
Wildfire Prediction System
Domain: Environmental ML / Climate Science
Live wildfire risk prediction for continental Portugal. A calibrated gradient-boosted ensemble fuses weather reanalysis, terrain, vegetation, and 25 years of fire history to forecast daily risk across a 1 km grid, served on a public map.
Archival Document Management AI
Domain: Document AI / NLP / Computer Vision
AI-driven document management and curation for a municipal archive. A five-stage Airflow pipeline runs OCR, layout analysis, named-entity recognition, LLM metadata enrichment, and multi-vector indexing, orchestrating 11 distinct ML models across GPU VRAM profiles.
Municipal MLOps & GPU Cluster
Domain: MLOps / Infrastructure
The production AI platform behind a Portuguese municipality. A GPU-backed Kubernetes cluster with GitOps automation and in-cluster CI/CD serves LLM inference and hosts the ML services its public-facing applications run on.
Real-time Conversational Avatar
Domain: Conversational AI / Speech
A real-time 3D talking avatar for municipal public service. Browser speech goes in, a custom LLM reasons, and the avatar answers out loud in the citizen's own language, streamed live over WebSocket from a GPU server.
Neural Forge is based in Santa Clara-a-Nova, in the municipality of Almodôvar, Baixo Alentejo. We work remotely with clients across Portugal and Europe, and on site with organisations in the Alentejo and the Algarve.
Much of our work is public-sector: wildfire risk for Portuguese territory, archival digitisation for a municipal archive, GPU infrastructure and citizen-facing conversational AI for local government. If you run a câmara municipal, an institute, or an industrial operation and need machine learning that actually reaches production, that is the work we do.
Almodôvar · Beja · Baixo Alentejo · Alentejo · Algarve · Portugal · Europe
Neural Forge is an applied machine learning and AI company based in Santa Clara-a-Nova, Almodôvar, in the Alentejo region of Portugal. We build end-to-end ML systems for environmental prediction, document understanding, conversational AI, and public-sector MLOps, covering everything from data collection and feature engineering through model training to deployed inference and the GPU infrastructure they run on. In practice that means we take on the whole problem rather than one slice of it: the ingestion pipelines, the models, the serving stack, the Kubernetes cluster underneath, and the interface people actually use. Four systems are currently in production, including a public wildfire risk map for continental Portugal and an eleven-model document AI pipeline for a municipal archive. We work in Portuguese and English, and most of our clients are Portuguese public bodies and institutes.
Neural Forge is based in Santa Clara-a-Nova, in the municipality of Almodôvar, district of Beja, Baixo Alentejo, Portugal. We serve clients throughout Portugal and Europe remotely, and work on site with organisations across the Alentejo and the Algarve. Being based in the interior of the Alentejo is deliberate rather than incidental: the wildfire risk, land, and heritage problems we work on are concentrated in exactly this kind of territory, and the institutions that own them (câmaras municipais, institutes, regional bodies) are often poorly served by Lisbon- and Porto-centric consultancies. Engagements typically run remote-first with periodic on-site work, and the infrastructure we build can be hosted on the client's own hardware.
Neural Forge's Tutela Ignis system combines satellite imagery, ERA5-Land weather reanalysis, NASA FIRMS active fire detections, terrain features, and 25 years of ICNF historical fire data into a calibrated XGBoost and LightGBM gradient-boosted ensemble that forecasts daily wildfire risk across a 1 km grid for continental Portugal. It is live at tutela.land. The pipeline engineers 63 features across weather, terrain, land cover, and vegetation moisture, and the ensemble is trained seasonally, because the drivers of fire in February are not the drivers in August. Raw model scores are not risk probabilities, so the system applies post-hoc probability calibration and a Bayesian prior shift at serving time, and is benchmarked against a stratified skill ceiling rather than against a baseline chosen to flatter it.
Yes. Public-sector work is the majority of what we do. Neural Forge builds and operates AI systems for Portuguese local government, including a GPU-backed k3s Kubernetes cluster with Flux CD GitOps automation and in-cluster CI/CD, an archival document AI pipeline running eleven ML models over a municipal archive, and a real-time multilingual conversational avatar for citizen services that answers in Portuguese, English, Spanish, or French. We understand what makes public-sector AI different from a startup prototype: procurement realities, archival standards such as ISAD(G), the need for systems that a small internal team can still operate in three years, and the fact that citizen data usually cannot leave the building. That last constraint is why we build on self-hosted GPU infrastructure rather than commercial AI APIs.
Machine learning research and model development, data engineering for multi-source pipelines, computer vision and OCR for document digitisation, LLM and RAG systems, and applied AI infrastructure: GPU clusters, Kubernetes, and MLOps. Services are delivered in Portuguese or English, remotely across Portugal and Europe or on site in the Alentejo. We are equally comfortable being handed a messy dataset and asked what is in it, being handed a trained model and asked to get it reliably into production, or being handed bare GPU hardware and asked to turn it into a cluster that serves inference. What we do not do is sell a fixed product: every engagement starts from the client's actual data and constraints, and every model has to earn its place through validation before it ships.
Yes. We run LLMs in production today: Qwen3 32B enriches archival metadata and Qwen3-VL 30B produces visual descriptions in our document AI pipeline, a LangGraph RAG chatbot answers questions over an entire municipal archive using multi-vector retrieval across Qdrant, and a custom Qwen3 30B persona drives a real-time conversational avatar for citizen services. Our approach to generative AI is retrieval-grounded rather than open-ended: models answer from a specific, curated corpus so their output can be traced back to a source document, which is what makes them usable in a public-service setting. We self-host on our own GPU hardware with Ollama rather than calling commercial AI APIs, so client data never leaves infrastructure the client controls.
Yes, and it is our default. Every ML system we have deployed runs on hardware we or the client own: a k3s Kubernetes cluster on RTX 5090 GPUs for municipal AI services, and an NVIDIA GB10 Grace Blackwell node serving the Tutela Ignis wildfire model. LLM inference runs on native Ollama, services are exposed securely through Cloudflare Tunnel with authenticated access rather than open ports, node provisioning is automated with Ansible, and the whole cluster is managed through Flux CD GitOps so its state is version-controlled and reproducible. For Portuguese public bodies this matters practically as well as legally: sensitive citizen and archival data stays in the country, on machines the institution controls, with no per-token dependency on an external AI provider.
No. Public bodies are the majority of our portfolio (municipalities, institutes, and national-scale environmental work), but the underlying capability is not specific to government. If you run an industrial operation, an agricultural or land-management business, or a company sitting on a dataset nobody has modelled yet, the work looks the same: understand the data, build the pipeline, train and validate the models, and deploy them somewhere they will keep running. We are a small team, so we take on a limited number of engagements and prefer problems where the data is genuinely difficult and the outcome is measurable. If you are not sure whether your problem is a machine learning problem at all, that is a reasonable first conversation to have.
We're always interested in problems where data meets real-world impact.
hello@neuralforge.pt