# Kroon Software House — Artificial intelligence systems in production > Brazilian artificial intelligence and MLOps infrastructure software house, based in Curitiba. > We build and operate computer vision systems, enterprise copilots, forecasting models and > machine learning platforms — with scope, deadline and SLA declared before the first > experiment. - Canonical origin: https://kroon.dev.br/en/ - Contact: projetos@kroon.dev.br · +55 41 9629-0481 · Curitiba, PR, Brazil - Languages: Portuguese, English and Spanish - Founded in 2017 · KROON SOFTWARE HOUSE S/A · CNPJ 12.831.549/0001-07 - Portuguese version: https://kroon.dev.br/ - Portuguese version in Markdown: https://kroon.dev.br/index.md?raw=1 ## Philosophy Technical rigor. Business judgment. AI systems reach production when the acceptance criterion is declared before the first experiment and measured on the client's data and infrastructure. ## What we do We build artificial intelligence systems that reach production and keep working: computer vision, copilots over documents, operations forecasting, the infrastructure that sustains them and the security that protects them. Every project starts from the operation's problem, not from the technology. ### AI development - **Computer Vision in Production** — RTSP cameras, video files or live streams start producing reliable events: counting, presence, PPE, queues, zone occupancy, line defects and access control. An inference pipeline with a defined accelerator (on-premises GPU or cloud); an event dashboard, search by period and alerts by webhook, email or WhatsApp; a target accuracy declared in the proposal and measured on validation cameras; deployment on your own server or in the cloud, with an operations manual. - **Enterprise Copilot over your documents** — search and answers with source citations over internal documents, standards, manuals and tickets, inside the client's environment, with per-user access control and an audit trail. Document ingestion and indexing, including scanned PDFs; an automatic evaluation layer (faithfulness, coverage and correct refusal); a web interface and API with usage logging and per-department permissions. - **Operations Forecasting and Optimization** — demand, inventory, predictive maintenance and routing, always measured against a published baseline. If the model does not beat the simple baseline, it does not go to production. ### Platform and operations - **MLOps Platforms and Pipelines** — CI/CD for models, version registry and reproducible training; inference on Kubernetes with autoscaling and an hourly cost ceiling; monitoring of model drift, latency and data quality in production; a quarterly review of cost per inference. - **Dedicated AI Squad** — a machine learning engineer, a platform engineer and a part-time tech lead, in two-week sprints with the scope agreed at the start of each cycle. Knowledge transfer and a guaranteed handover. ### Security - **AI Application and Infrastructure Security** — an attack-surface map, with a review of access, secrets and dependencies; hardening of cluster, inference and pipelines with least-privilege policy; anomaly detection in logs; an incident response plan rehearsed in a simulation. - **Penetration Testing for Applications, APIs and Models** — scope and rules of engagement defined before the first test; manual and automated testing; AI-specific cases (prompt injection, inference leakage, endpoint abuse); a report with step-by-step reproduction, severity and retesting. We also work with machine learning and deep learning, natural language processing, autonomous agents, cloud architecture and Kubernetes. ## The Kroon method (five named stages) 1. **Technical assessment** — 30 minutes with an engineer, free of charge, with a written opinion. 2. **Scope, investment and acceptance criteria** — declared before any code is written. 3. **Proof of value with your data** — week 3; the agreed metric is measured on the client's data and infrastructure. If it is not met and there is no declared technical path, the project ends there and the client keeps the code, the experiments and the report. 4. **Production and acceptance** — go-live in 6 to 8 weeks in most scopes. 5. **Support and evolution** — continuous operation, with a guaranteed handover to the client's team whenever they want it (manual, videos and 30 days of side-by-side support). ## Technology Computer vision · modeling and research · data and language · platform and inference · cloud and observability. ## Frequently asked questions **Do you use our data?** Yes, and only under a data processing agreement signed before the first access. We follow LGPD and GDPR strictly, from collection to disposal. The client chooses the arrangement: training inside their environment, prior anonymization, or a dedicated cloud account under their control. No client data ever enters a model that serves another client. **What if the model does not reach the agreed acceptance criterion?** That is precisely what the proof of value in week 3 is for. We measure the agreed metric on the client's data and infrastructure before building the rest. If the target is not met and there is no declared technical path, the project ends there, and the client keeps the code, the experiments and the measurement report. **How long until the first usable result?** Three weeks to a proof of value measured on the client's data, and six to eight weeks to go live in most computer vision and enterprise copilot scopes. Forecasting projects take longer because of the baseline comparison cycle: generally eight weeks. **Do you replace our technology team?** No. We work in the team's repository, with code review in both directions, weekly pairing and mandatory documentation. Every project includes a guaranteed handover. **Do you work with companies outside Brazil?** Yes. We work in Portuguese, English and Spanish and cover meetings from the Lisbon time zone to the west coast of the United States. For data with a residency requirement, we keep training and inference in the region the client chooses. **Do you do a free proof of concept?** No. A free proof of concept usually means a model rushed out on data that does not represent production. The 30-minute assessment and the written opinion are free; the proof of value is the first stage of the project, with its own scope and acceptance criteria, not a courtesy sample. ## Contact Talk to an engineer: form at https://kroon.dev.br/en/#contato · projetos@kroon.dev.br · +55 41 9629-0481 (BRT business hours). Address: Rua Padre Anchieta, 2540, 15th floor — Curitiba, PR, 80730-000, Brazil.