Each project starts with a real problem and a dataset. We build the pipeline, train the models, deploy the infrastructure, and ship the interface.
01Live
Tutela
Wildfire Platform: Ignis and Limes
Domain: Environmental ML / Remote Sensing
Tutela is Neural Forge's environmental prediction platform. Ignis, its founding module, is live at tutela.land: a multi-source pipeline ingests ERA5-Land reanalysis, IPMA weather, NASA FIRMS active detections, ICNF fire records, Copernicus imagery, DEM terrain, land cover, and vegetation moisture into a seasonal gradient-boosted ensemble with post-hoc probability calibration and a Bayesian prior shift at serving, forecasting daily fire risk for 88,527 one-kilometre cells across continental Portugal. Validation is prospective and public: a live ledger tracks every forecast against real ignitions, and at matched ignition coverage the system flags 1.7 to 3.4 times less area than the top alert levels of the official systems. The methodology is published as an open-access preprint on Zenodo, and the research code and daily prediction archive are open on GitHub. Limes, the second module, works on the ground side of the same problem: the legally required fuel-management strips along roads, power lines and around buildings, which a small municipal team cannot inspect on foot every season. Limes crosses the official strip network with the cadastre, measures the vegetation standing in every parcel from the 2024 national LiDAR survey at 50 cm, and ranks parcels by hectares of tall vegetation, a transparent number anyone can recheck. Weekly satellite comparison adds a change signal that only nudges the ranking and abstains where a strip is too narrow to be seen, and burned area is confirmed from the EU fire-mapping service. The output is a ranked worklist per freguesia, an evidence sheet per parcel and an export for the municipality's own GIS: situations to verify, never verdicts, and no substitute for the technician's assessment. A first pilot of Limes is under evaluation with the Municipal Civil Protection Service of Vila Franca de Xira, within a research project of CIAITeX, the municipal centre for artificial intelligence, innovation, and IT (2026). The platform runs on a single self-hosted edge node with nightly disaster-recovery backups and hourly external health alerting.
Ignis: daily risk on a 1 km grid, public map and prospective validation ledger at tutela.land
Calibrated seasonal ensemble: 1.7 to 3.4x less area flagged than official top alert levels, at matched coverage
Open-access methods preprint on Zenodo; research code and prediction archive on GitHub
Limes: fuel-management strips ranked from national LiDAR and weekly satellite change; situations to verify, never verdicts
Scriba (Latin: the scribe) is the document AI behind the CIAITeX portal of a municipal archive, a collection of roughly 115,000 descriptive records reaching back to 1834. A 5-DAG Airflow pipeline ingests descriptive records and digital objects from the archive management system, runs page layout analysis and OCR with a vision-language transcription path for handwritten material, extracts named entities, classifies personal-data sensitivity into four GDPR-aligned levels, and enriches ISAD(G) metadata with a self-hosted LLM, labelling every AI-generated field as such. Text, visual, and multi-vector embeddings power hybrid semantic retrieval over Qdrant with reciprocal-rank fusion and reranking, and a LangGraph RAG chatbot answers questions grounded in retrieved archival evidence. Eleven open-weight models are orchestrated VRAM-aware on a single 32 GB GPU, and every write back into the archival catalogue is human-reviewed and gated. Deployed and in pilot operation, processing historical record series end-to-end.
11 open-weight models, VRAM-aware on a single 32 GB GPU
Hybrid multi-vector retrieval with rank fusion and reranking
ISAD(G) and GDPR-aligned enrichment, AI output always labelled
Human-reviewed, gated write-back into the archival catalogue
Sovereign AI means the models, the data, and the hardware all answer to the institution that owns them. Neural Forge designs, builds, and operates on-premise GPU infrastructure: self-hosted language-model serving with no per-token dependency on outside AI providers, and machine learning services that keep sensitive data inside the organisation. The entire state of a cluster is declared in code, so every change is reviewed, auditable, and reproducible; provisioning new nodes is automated; services are reached securely without opening ports to the internet; and the documentation ships with the cluster. Built so a small internal team can still operate it in three years. Already in production in the Portuguese public sector, where it runs the document AI and conversational services of local government.
Self-hosted LLM serving on client-owned GPUs
Cluster state declared in code: reviewed, auditable, reproducible
Secure remote access with no ports open to the internet
Built for a small internal team to operate for years
0data off-site
3+services
99%availability
04Live
Arauto
Conversational AI Avatar
Domain: Conversational AI / Speech
Arauto (Portuguese for herald, the figure who spoke for an institution to its people) is Neural Forge's conversational avatar product. It learns your organisation, whether that is a company, a public institution, or a place, from a knowledge base curated with you. A visitor speaks through the browser; Arauto listens, reasons over that knowledge, and answers out loud in the visitor's own language, with automatic detection across six languages (Portuguese, English, Spanish, French, German, and Italian) and sentence-by-sentence audio streaming in chunks of roughly 100 ms, so the conversation feels immediate. It runs entirely on self-hosted infrastructure: speech recognition, the language model, and voice synthesis all operate on hardware the organisation controls, and no conversation ever reaches an outside AI provider. Explicit refusal rules keep it inside its brief instead of improvising. Already in production in Portuguese local government, and its system design is described in a peer-reviewed paper accepted at TEEM 2026.
Real-time speech-to-speech in the browser
Answers only from your organisation's curated knowledge
Fully self-hosted: conversations never reach an outside AI provider
Peer-reviewed system design, accepted at TEEM 2026
6languages
~100msaudio chunks
TEEM 2026paper
05Live
Faber
Grounded Site Assistant
Domain: Conversational AI / Websites
Faber (Latin: the smith, the one who works the forge) is Neural Forge's grounded site assistant. Everything it knows is in one plain-text facts file the organisation controls; when a question falls outside it, Faber says so and gives the organisation's contact instead of improvising a plausible answer. It replies in the visitor's own language, native European Portuguese included, cites the site's pages and email as real links, and never quotes prices, timelines, or commitments. It can run hosted, on a small language model processed inside the EU behind a fixed monthly budget with a hard cap, or self-hosted on GPUs the organisation controls, on the same infrastructure Neural Forge already operates in Portuguese local government, so no conversation reaches an outside AI provider. Nothing is stored: the conversation lives in the browser tab. Faber is in production on neuralforge.pt, where every answer on this page was recorded.
Answers only from a facts file you control; says so when it does not know
Replies in the visitor's language, native pt-PT
Hosted in the EU with a hard budget cap, or self-hosted on your GPUs
In production on this site: try it at the bottom right
Outrora (Portuguese for the time that once was) is Neural Forge's virtual visits product. A place is captured in 360° on location, each image oriented from the position of the sun, and a talking 3D guide tells the story of every stop in narrations written with the people who know the place and recorded in Portuguese and English. The same visit works three ways: at home as one continuous 360° walk in the browser; on site, where GPS, compass and camera recognition put the guide at the door of each place on the visitor's phone; and, in testing, in a virtual reality headset. The guide does not improvise: he says only what was written for each stop. Visits are published on the customer's own domain, in the customer's brand, with Neural Forge in the signature only, and grow over time without going down. The first visit, Santa Clara-a-Nova in Almodôvar, went live in September 2026.
Photographed in 360° on location, as the place is today
A talking guide, at home in the browser or on site on the phone
Published on the customer's domain, in the customer's brand
Content stays with the customer; nothing to install, no tracking