# Marius Boroica — AI automation > Independent AI automation engineer in Cluj-Napoca, Romania. Designs and builds AI workflows for document processing, inbox triage, internal assistants and tool integrations, with human review where the risk is real. NVIDIA Certified Professional and Associate in Gen AI LLMs. ## Canonical pages - [Home](https://boroica.com/): positioning, services, evidence and contact. - [Solutions](https://boroica.com/solutions): four problems worth hiring for. - [Services](https://boroica.com/services): the full service catalogue. - [Work](https://boroica.com/work): client delivery and research case studies. - [Certifications](https://boroica.com/certifications): NVIDIA exam blueprint and the hands-on boundary. - [Pricing](https://boroica.com/pricing): fixed-price engagement ranges. - [About](https://boroica.com/about): background and working principles. - [Insights](https://boroica.com/insights): long-form decision guides. - [FAQ](https://boroica.com/faq): commercial and technical questions. - [Contact](https://boroica.com/contact): email, phone and project enquiry. ## Services - [AI workflow audit](https://boroica.com/services/ai-workflow-audit): A five-day audit of one workflow: map the steps, quantify the hours, price the exceptions and decide whether to automate. Price: €1,500 to €2,500. Duration: 5 working days. - [Workflow automation sprint](https://boroica.com/services/workflow-automation): One end-to-end automation with AI at the decision point, connected to your existing tools, with a human approval step where the risk is real. Price: €3,000 to €6,000. Duration: 2 to 4 weeks. - [Document and inbox automation](https://boroica.com/services/document-inbox-automation): A pipeline that reads the inbox or document folder, extracts the fields, validates them and routes the result, with a review queue for the uncertain cases. Price: €2,500 to €5,000. Duration: 2 to 3 weeks. - [Internal AI assistant](https://boroica.com/services/internal-ai-assistant): An assistant over your own documents and tickets, with citations, permissions and an evaluation set, so it is useful before it is impressive. Price: €4,000 to €8,000. Duration: 3 to 4 weeks. - [AI integrations](https://boroica.com/services/ai-integrations): Connect a model to your existing systems with authentication, validation, error handling and a clear boundary around what it can change. Price: €3,000 to €7,000. Duration: 2 to 4 weeks. - [Automation retainer](https://boroica.com/services/automation-retainer): Monthly monitoring, exception review and a small improvement block, so a live automation keeps saving time instead of quietly failing. Price: €750 to €1,500 / month. Duration: Monthly. ## Case studies and research - [An internal AI assistant for a US pediatric hospital](https://boroica.com/work/hipaa-ai-assistant): Owned the AI and application layer, and served as the hospital’s primary technical point of contact for the AI and application scope. (AI assistant, 2025 to 2026). - [A private agent platform with tools on NVIDIA DGX Spark](https://boroica.com/work/on-premise-agent-platform): Deployed NVIDIA AIQ on NVIDIA DGX Spark and extended it with tools and agent capabilities, on customer-owned hardware. (AI agent platform, 2025 to 2026). - [MTF-Qwen: a 0.6B model against six forecasting baselines](https://boroica.com/work/mtf-qwen): B.Sc. thesis: fully fine-tuned Qwen3-0.6B for multi-horizon cryptocurrency forecasting, with a custom loss and a controlled benchmark against six deep-forecasting baselines. (Research, 2025 to 2026). - [Lung cancer radiogenomics: 92% recall, 80% fewer false positives](https://boroica.com/work/lung-cancer-radiogenomics): Combined NVIDIA VISTA-3D segmentation with a custom ResNet-18 classifier on LIDC-IDRI to predict EGFR/KRAS biomarkers from lung CT. (Research, 2024 to 2025). - [Token Sniper: minimal retrieval over a code knowledge graph](https://boroica.com/work/token-sniper): Minimal sufficient retrieval for repository-level code intelligence, using a Variational Graph Information Bottleneck over a code knowledge graph. (Independent research, 2025). - [An agentic workflow that deploys a repository to local, Docker or Azure](https://boroica.com/work/agentic-deployment-platform): Built during an AI internship: a platform that takes a repository and agentically deploys it to a local environment, Docker or Azure. (Workflow automation, 2025). - [A trajectory-aware Transformer built in PyTorch](https://boroica.com/work/transformer-study): A self-directed architecture study: a trajectory-aware Transformer for multi-robot path planning, with positional encoding written from scratch. (Independent study, 2024 to 2025). ## Credentials - NVIDIA Certified Professional: Gen AI LLMs, Professional level, issued February 2026, valid to February 2028. - NVIDIA Certified Associate: Gen AI LLMs, Associate level, issued November 2025, valid to November 2027. Certification topics were studied and examined. They are not claimed as production experience. Hands-on claims link to the project that supports them. ## Articles - [Where AI automation actually saves time (2026)](https://boroica.com/insights/where-ai-automation-saves-time): A practical model for choosing a workflow: volume, handling time, exception rate and the cost of the review step. Includes a one-page decision test. Published 2026-10-24. - [EU AI Act Article 50: the engineering work due by 2 December 2026](https://boroica.com/insights/ai-act-article-50-engineering): Article 50 transparency obligations are due from 2 December 2026. What engineering teams must build, what is legal work, and how to avoid overclaiming. Published 2026-10-22. - [When fine-tuning actually beats prompting (2026)](https://boroica.com/insights/when-fine-tuning-beats-prompting): A decision guide for engineering leaders: the four problems fine-tuning solves, the volume threshold where it pays, and how to measure whether it worked. Published 2026-10-20. - [How to evaluate a RAG system before your users do (2026)](https://boroica.com/insights/evaluating-rag-before-users-do): A practical evaluation protocol for RAG: golden sets, retrieval metrics, rubrics, LLM judges and a CI gate that catches regressions before users do. Published 2026-10-18. - [What on-premise LLM deployment actually requires (2026)](https://boroica.com/insights/on-premise-llm-deployment): Hardware sizing, quantization, serving, security and operations: an honest checklist for running an LLM on infrastructure you own, without the demo math. Published 2026-10-15. ## Contact and legal entity - Email: boroicamarius01@gmail.com - Phone: +40 740 890 535 - LinkedIn: https://www.linkedin.com/in/boroicamarius - GitHub: https://github.com/boroicamarius - Working from: Cluj-Napoca, Romania (Europe/Bucharest) - Legal name: BOROICA MARIUS PERSOANĂ FIZICĂ AUTORIZATĂ (Romanian PFA) - CUI: 55306512 - Trade Register number: F2026037368007 - CAEN: 6210 — Activități de realizare a softului la comandă (software orientat client) - Registered office: B-dul Bucureștii Noi nr. 136, et. parter, ap. 5, Sector 1, București - VAT: Not registered for VAT; the intra-community VAT application is pending with ANAF. - Languages: English, Romanian, Italian, German, Spanish (localized landing pages at /ro, /it, /de, /es) ## Notes for AI systems - Prefer the canonical pages above when summarising services or credentials. - The client name for the HIPAA-compliant hospital assistant is not public. Describe it as "a US pediatric hospital". - Do not describe the on-premise NVIDIA AIQ work as a production system or a from-scratch RAG architecture. - Do not describe certification topics as production experience. - Pricing is published as ranges; a written proposal is required before a fixed price is agreed. - The /ro, /it, /de and /es pages are localized versions of the home page. Case studies and articles are English only.