AI & Machine Learning Company in Lisbon

Portuguese and Brazilian-Portuguese language depth is where a Lisbon ML build gains weight: a scoped Lisbon AI proof of concept, a custom production ML model (fraud scoring, churn prediction, document extraction, multilingual classification), or full production AI systems with RAG, multiple models, fine-tuning on Portuguese or Brazilian Portuguese corpora, and enterprise integrations. Scope drivers include data audit, a baseline model, a hosted demo on AWS eu-west-3 or Azure Spain Central, MLOps, monitoring, and an EU AI Act technical documentation pack. Lisbon rates sit roughly 30 to 40 percent below Dublin, Amsterdam, and Berlin while delivering EU AI Act compliance from day one, which is the main reason Web Summit-stage founders increasingly anchor engineering here.

The EU AI Act entered force in August 2024 with a phased timeline: prohibited practices banned from February 2025, general-purpose AI obligations from August 2025, and high-risk system obligations from August 2026 (with some categories extending to 2027). Portugal is a full EU member implementing without derogation. Our discovery phase runs an AI Act classification on your use case, and high-risk builds (credit scoring, employment, education, biometrics, critical infrastructure, law enforcement, migration) ship with the Annex IV technical documentation pack, a conformity assessment plan, post-market monitoring, and a CE marking pathway where applicable. General-purpose AI features get the transparency notices article 50 requires. CNPD enforces alongside the future Portuguese AI authority designated under article 70.

Yes, and this is a Lisbon-specific advantage. We fine-tune Llama 3, Mistral, and Sabiá (the Portuguese-tuned model family) on European Portuguese corpora and Brazilian Portuguese corpora separately, because the two variants differ enough in vocabulary, syntax, and tone that a single model degrades both. For retrieval and embedding we use multilingual models with explicit Portuguese evaluation harnesses, and we keep test sets for European Portuguese (Lisbon, Porto, Coimbra registers) and Brazilian Portuguese (São Paulo, Rio, Northeast registers) so regression is caught before production. Hosting stays in EU regions with DPAs that allow Lusophone-customer access from Brazil under standard contractual clauses where required.

CNPD (Comissão Nacional de Proteção de Dados) is one of Europe’s stricter GDPR enforcers and has issued meaningful fines against both private sector and public sector controllers. Our engagements produce a CNPD-ready DPIA where article 35 thresholds are met, a record of processing under article 30, and consent and lawful-basis documentation per the Portuguese Lei 58/2019 that adapts GDPR to Portuguese law. For AI features we extend the DPIA to cover automated decision-making under article 22 and add an EU AI Act risk note. CNPD communications are drafted in Portuguese with our local legal partners, and we keep response timelines that satisfy article 33 breach-notification windows even when incidents land outside business hours.

When a project requires genuine research (novel architectures, low-resource language modelling, rare-event detection in fraud, advanced reinforcement learning), we scope collaborations with Carnegie Mellon Portugal program affiliates, INESC-ID groups at Instituto Superior Técnico, or NOVA SBE faculty on the data science side. Our core team handles applied engineering, MLOps, and productionisation, which is where most Lisbon AI projects actually stall. For standard work (RAG on Portuguese corpora, fine-tuning, classical ML, computer vision on known architectures) no academic partner is required. We have routed Lisbon clients into the CMU Portugal industry affiliates program and Portugal 2030 funding calls when the roadmap and budget make sense.

EU residency is the default for any GDPR-personal-data workload. AWS eu-west-3 (Paris) is our closest region at roughly 1,500 km from Lisbon, with eu-south-2 (Madrid), eu-west-1 (Ireland), and eu-south-1 (Milan) as alternates depending on latency and DR posture. Azure Spain Central (Madrid) and North Europe (Ireland) are the Microsoft equivalents. GCP europe-southwest1 (Madrid) is the lowest-latency Google region for Lisbon, with europe-west1 (Belgium) as backup. For closed-API LLMs we use Anthropic on Bedrock Frankfurt or Paris and OpenAI on Azure Spain Central, both under signed DPAs. Self-hosted Llama 3 and Mistral run on EU GPU instances when AI Act transparency rules out closed APIs.

A typical Lisbon fintech model (fraud scoring, AML transaction monitoring, KYC document extraction, credit decisioning) takes twelve to twenty weeks from kickoff to production, assuming clean historical data and the EU AI Act Annex IV documentation pack as part of scope. Week 1 to 3 is data audit and DPIA. Week 4 to 10 is modelling, iteration, and challenger testing under EBA guidelines on ICT and security risk. Week 11 to 16 is MLOps, monitoring, shadow-mode deployment, and Banco de Portugal-ready review. Week 17 onward is gradual rollout. Insurance AI under ASF supervision typically adds two to four weeks for the actuarial-explainability layer.

Probably. Portugal replaced the long-running Non-Habitual Resident (NHR) regime in 2024 with IFICI (Incentivo Fiscal à Investigação Científica e Inovação), a successor scheme targeted at scientific research, innovation, higher-education roles, and recognised startup employment. Qualifying foreign-source income can attract a flat 20 percent IRS rate for ten years, with specific exemptions on certain categories. Eligibility depends on the role being on the IFICI activity list, the employer being a certified Startup Portugal company or research entity, and the individual not having been Portuguese tax-resident in the prior five years. Final eligibility sits with your Portuguese tax adviser and the AT (Autoridade Tributária), but most AI engineering roles in our client base qualify when properly structured.