Brazilian AI infrastructure
for regulated sectors.
BeansTech operates 206 production AI services across 11 regulated verticals. 67M court decisions indexed. GCP Digital Leader. Migrating to Alibaba Cloud with Qwen 3.8 as primary LLM. São Paulo, Brazil.
AI infrastructure for Brazil's most regulated sectors.
LegalTech
67M court decisions indexed. RAGJur anti-hallucination protocol (INPI-filed). A fabricated citation can cost a Brazilian lawyer their OAB license — we solved this.
HealthTech
AI medical imaging analysis (94% accuracy). Automated SOAP clinical notes. Einstein hospital requested a demo. CFM 2.454 + LGPD compliant.
FinTech / RegTech
AML compliance. Credit scoring. Tax analysis. BACEN + CVM regulation. Qwen numerical reasoning handles Brazilian financial regulatory complexity.
ConTech + Diálogo
11-module construction suite. 25 functions. Jobsite photo analysis. RFI drafting. Brazil's construction market: R$ 800B/year.
PropTech
Property valuation. Rental management. Deed OCR. Brazil has 80M+ urban properties — massive document-intelligence demand.
ESG / Land Regularization
30-agent workflow. Nationwide land regularization. Qwen as the bulk reasoning engine — cost-optimized, high throughput.
Matheus Ximenes Feijão Guimarães
OAB/SP 524.129 · Master's candidate, Law & New Technologies
Lawyer, 12 years at Brazil's Superior Military Tribunal (STM). 7 years as legal advisor to Admirals of the Fleet — including the current STM President (Almirante de Esquadra Leonardo Puntel). 5 postgraduate degrees.
GCP Digital Leader · Google for Startups · Claude Partner. Apache 2.0 open-source contributor — AI security for the Brazilian Judiciary (OWASP LLM01:2025). 11 categories of prompt-injection detection.
Deep respect for China's long-term vision and technical excellence — built over years of daily interaction with the Chinese community in Brazil. Office: Pátio Mall Zone, 8th floor — same tower as the Bank of China's Brazil headquarters.
- ◆ Master's candidate, Law & New Technologies
- ◆ 5 postgrad degrees (Cloud, LGPD, Criminal, Growth+AI, EdTech)
- ◆ 12 years STM — 7 as advisor to Fleet Admirals
- ◆ GCP Digital Leader · Google for Startups
- ◆ Claude Partner (Anthropic)
- ◆ Apache 2.0 open-source · AI security
- ◆ Pátio Mall Zone, São Paulo
Chinese models are not alternatives — they are state of the art.
After 90,000+ chat turns with Claude and deep production experience with multiple LLMs, our conclusion is objective: Chinese AI models — particularly Qwen and its specialized variants — are the best fit for Brazilian regulated workloads. Not because they're cheap. Because they're the best for our use cases.
Qwen 3.8 (通义千问)
Strong multilingual reasoning (PT-BR legal/medical/financial), low hallucination baseline, long-context capabilities. The ideal engine for RAGJur's 67M legal documents and e-arbitragem.ai's arbitration sentence drafting.
Qwen-VL
Document OCR for legal petitions, medical prescriptions, property deeds, construction blueprints, and ID cards. Powers our OCR pipelines across LegalTech, HealthTech, PropTech, and ConTech.
Qwen Medical
Radiograph analysis with 94% accuracy. Clinical SOAP note generation. Patient triage. Brazilian Portuguese, LGPD-compliant. Einstein hospital requested a demo of this stack.
Qwen Embeddings
Powers semantic search across 67M court decisions, medical records, and financial documents. The backbone of RAGJur's anti-hallucination verification layer.
GLM-5 / GLM-5 Zen
Already running in production at useco2.ai as the cost-optimized bulk tier ($0.50/laudo, 0.9s). The only model that satisfied the cost + low-hallucination + PT-BR legal reasoning triangle for RAGJur.
DeepSeek R1
ESG scoring and financial calculations. Strong numerical reasoning for Brazil's complex BACEN/CVM regulatory environment. Part of our multi-model production stack.
We benchmark Chinese models head-to-head against Western alternatives in real production workloads — not synthetic tests. The results: Qwen 3.8 matches or exceeds Claude and GPT on Brazilian Portuguese legal/medical reasoning, at a fraction of the cost. DeepSeek R1 leads on numerical tasks. GLM-5 is unmatched on cost-per-quality for bulk inference. This is not ideology — it's production data.
We partner with global leaders.
Alibaba Cloud
Full migration in 30 days. Qwen 3.8 as primary LLM. 10 data center sites analyzed. Visit: alibaba.beanstech.com.br
Z.ai
GLM-5 already in production at useco2.ai. The only model optimizing RAGJur. Visit: z.beanstech.com.br
Google Cloud + Anthropic
GCP Digital Leader. Google for Startups. Claude Partner. 90,000+ Claude chat turns — deep LLM infrastructure experience.
Built in São Paulo. For Latin America.
Qwen-ready · Alibaba Cloud · Google Cloud · Anthropic
+55 (11) 92507-9058 · São Paulo, Brazil · Pátio Mall Zone