01
Agents in production
Task-specific agents wired into the tools you already run, with access control, logging and a defined path for the cases they should not handle alone.
Applied artificial intelligence · São Paulo
AI systems that enter the operation, and stay.
NASCIA AI designs, builds and runs agents, evaluation harnesses and integrations inside the processes your organisation already uses every day. Not a demonstration: a system your own team can operate, audit and extend.
01. Define
Name the task, the owner and the measure of success before any model is chosen.
02. Instrument
Build the evaluation harness first, so improvement can be told from luck.
03. Integrate
Ship into the existing workflow, with fallbacks, logging and access control.
04. Hand over
Leave the runbook, the tests and the team that can operate it without us.
The same four steps, in the same order, on every engagement.
What we build
We build the part that has to work.
An agent is easy to demonstrate and hard to keep running. The difference is not the model. It is everything around it: the data it is allowed to see, the evaluation that tells you when it degrades, the fallback when it fails and the person who owns the outcome.
02
Evaluation harnesses
A test bed that tells a real improvement from a lucky run, so a model or prompt change never ships on a hunch.
Model
ERP
CRM
Ticketing
Finance
Documents
03
Integration and automation
The plumbing between the model and your systems of record: ERP, CRM, ticketing, finance and documents, built to be audited.
04
AI training for organisations
Practical programmes for companies, schools and foundations, taught by someone who builds these systems as well as writes about them.
In production
Shipped, and still running.
Four public products and one confidential enterprise system. Each card states one fact we can verify, with its source and date, and links to the full case study.
Education
AI Teachers
15,000+ · educators a month, as stated on ai-teachers.pro and on the founder's site. Source: ai-teachers.pro and mrnascimento.com, fact ai-teachers-01, 24 September 2026.
Read the case study
Finance
Portfolio Horizon
415 · tests passed in the adversarial suite behind the fail closed output boundary, 11 files, one run. Source: Project claims allowlist EXT-B03, fact portfolio-horizon-12, 31 July 2026.
Read the case study
Education
Careers Studio
12 · stages in the AI orchestration that maps each topic. Source: careers-e4c8e.web.app landing page (careers-studio-06), 24 September 2026.
Read the case study
Education
Mr Nascimento Academy
480 · IGCSE 0607 PDFs in the protected archive. Source: mrnascimento.academy home page (mr-nascimento-academy-11), 24 September 2026.
Read the case study
Confidential client
Enterprise finance agents
For a privately held group: the first line of the monthly income statement (DRE) produced automatically from raw ERP data, validated against six months of historical data and accepted in writing by the executive sponsor.
Delivered through more than ten versioned releases, with more than 60 pages of documentation.
Read the case studyIllustration with simulated data. The scores are generated for this demonstration and do not come from a client system.
Improvement or luck?
Two versions of a model are scored on the same evaluation set. Move the slider and watch what the sample size does to the answer.
Number of evaluation examples
At 100 examples the intervals overlap. This picture alone cannot establish whether the versions differ. Overlap is not evidence that their performance is equivalent.
- Version A
- Accuracy: 82.0% (82/100)95 per cent confidence interval (Wilson): [73.3%; 88.3%]Interval width: 15.0
- Version B
- Accuracy: 86.0% (86/100)95 per cent confidence interval (Wilson): [77.9%; 91.5%]Interval width: 13.6
Each score has a Wilson 95 per cent confidence interval. The display illustrates uncertainty, not a significance test. Comparing models on the same examples requires a paired analysis of their outcomes, which these simulated aggregate scores do not contain.
How we work
Four steps, in this order, every time.
The order is the method. Nothing is built before it is defined, nothing ships before it is measured, and nothing is handed over without the people who can keep it running.
01. Define
Name the task, the owner and the measure of success before any model is chosen.
Deliverables
- The definition: the task, its owner, the measure of success and the data the system may see.
02. Instrument
Build the evaluation harness first, so improvement can be told from luck.
Deliverables
- The evaluation harness: a test set from real cases, metrics, baselines and a report that runs on every change.
03. Integrate
Ship into the existing workflow, with fallbacks, logging and access control.
Deliverables
- The integrated system, inside the tools you already use, with logging, fallbacks and access control.
04. Hand over
Leave the runbook, the tests and the team that can operate it without us.
Deliverables
- The runbook, the tests and a team trained to operate the system without us.
A system nobody uses is a slide.
The measure of the work is not the demonstration. It is whether the system is still running, still measured and still owned by your team after we have left.
Training
Taught by the person who builds the systems.
Programmes for companies, schools and foundations, in Portuguese and English, from executive briefings to hands-on workshops for engineering teams and training for teachers.
See the training programmesCompanies
Leadership and engineering teams that have to decide what to build, what to buy and how to measure it.
Schools
Teachers and school leaders who need a working AI policy and classroom practice that holds up.
Foundations and public bodies
Teams that fund or run education and public services and need to judge AI proposals with evidence.
Founder
Built by someone who ships the systems and teaches the people.
Roney Lima do Nascimento
Founder, NASCIA AI
Author, mathematician, AI solutions developer, applied AI researcher, speaker and teacher.
Author of Generative AI for Teachers: A Practical Guide to Educational Technology (2025) and of Inteligência Artificial Generativa para Professores (Editora Dialética, 2026).
Mathematician. PhD candidate in Pure Mathematics at IME-USP, University of São Paulo, with the qualifying examination passed in 2026.
AI solutions developer who builds intelligent systems for large organisations and leads their AI implementation from first definition to hand over.
Applied AI researcher. Peer reviewer for Springer's Education and Information Technologies and for the International Conference on Educational Data Mining.
Keynote speaker at ICAILY 2026 in Cape Town. Speaker at the Summit de Inteligência Artificial Brasil 2026 in Joinville and at MiningTech South America 2026 in São Paulo.
Trains teachers in schools across Brazil and abroad, including AI literacy training for more than 50 teachers in one school programme, written up for ISTE+ASCD Educational Leadership. Has taught mathematics for more than fifteen years, including in international schools.
Credentials
- Claude Certified Architect, Professional (Anthropic, 2026)
- Claude Certified Architect, Foundations (Anthropic, 2026)
- Google Generative AI Leader
- IBM Generative AI Engineering Professional Certificate (2026)
- MBA, Ibmec
- PhD candidate, IME-USP
Books
- Folha de S.Paulo
- EUobserver
- GovInsider
- Business Day
- HEPI
- The AI Journal
- eSchool News
- Congresso em Foco
- Nexo Políticas Públicas
- La Prensa (Panamá)
- El País (Uruguay)
- RH Pra Você
- ISTE+ASCD Educational Leadership
Research and writing
How we think about evaluation, in public.
A forensic study of answer leakage in agentic code repair, and opinion pieces on AI evaluation, education and public policy in the Brazilian and international press.
Study
When Execution Feedback Reveals the Expected Output: A Forensic Study of Answer Leakage in Agentic Code Repair
Zenodo, author preprint, not peer reviewed. DOI 10.5281/zenodo.21286873,
A preprint by NASCIA's founder audits three of his own experiments, finds the expected output leaking through execution feedback, and sets out the controls a credible result needs.
ReadVerification debt is the hidden cost of enterprise AI
The AI Journal, , English
The human-review queue is the next digital divide
GovInsider, , English
AI chatbots need to work in all EU languages, not just English
EUobserver, , English
Questions
What people ask before the first conversation.
How does an engagement work?
In four steps, always in the same order: define, instrument, integrate, hand over. We sit with your team, map the workflow, ship a working system into the tools you already use and measure it in production. Scope and duration depend on the process; the first conversation is enough to size it.
Do you name your clients?
No. Enterprise work is described without names, as a privately held group or as large organisations, unless a client asks to be named. The case studies on this site show only what we can verify and are allowed to publish.
How do you handle our data and the LGPD?
Each system is defined with the data it is allowed to see, and nothing more, before any model is chosen. Access control, logging and fallbacks are part of the build, not an add-on. You remain the controller of your data; we work under the Brazilian General Data Protection Law (LGPD) and under your own policies, and the definition step records where data may and may not go.
Which languages do you work in?
Delivery happens in Portuguese and English: workshops, documentation, runbooks and code. This site is also published in Spanish.
What do you not do?
We do not sell slides, and we do not resell models or licences. We do not build a demonstration that nobody will operate. If a process does not need an agent, we say so at the definition step.
How do we start?
Write to us with the process that keeps breaking, who owns it, the systems it touches and what a good result would look like. One conversation is usually enough to tell whether an agent belongs there and what it would take to keep it running.
Bring us the process that keeps breaking.
One conversation is usually enough to tell whether an agent belongs there, what it should be measured against, and what it would take to keep it running.