01
About Guidevit
AI agents that talk to real customers need a layer that keeps them trustworthy. That layer is my work.
Most recently I was the founding AI engineer at DealsmartAI, where I owned the AI stack end to end: the agent loop and tool design, MCP, memory, context engineering, evals, tracing, and the compliance layer that governed every outbound message.
Read more
Two convictions shape how I build. Context is assembled, not dumped — most agent failures are context problems wearing a model costume. And compliance is architecture, not a paragraph in a prompt: a rule the business can't afford to break belongs in code the model cannot bypass.
Evals are what make the rest possible — baseline before, delta after, suites gating CI. When an agent misbehaves there's no stack trace for a bad decision; the answer is in the spans. So the failure classes unique to agentic systems get their own instrumentation: denied tool calls, silent terminations, safety-net fires.
The same principle scales to agents that write code. I built the multi-agent harness that shipped the product — subagents fan out, findings are adversarially verified, convergence is computed rather than self-declared. The median pull request went from opened to merged in hours, not days.
Before engineering, 18 months as a Product Owner. Which is why I take problems, not specifications.
Today I'm building Nicio (nicio.com.br), my own AI product for trademark lawyers in Brazil. Open to remote AI Engineer roles in US and Latin American time zones.
Python · Claude Agent SDK · MCP · FastAPI · Braintrust · PostgreSQL · Qdrant · Neo4j
Professional profile
- Location
- Curitiba, Brazil
- Time zone
- America/Sao_Paulo
- Work mode
- Remote
- Availability
- Available Now
- AI experience
- 3+ years
- Industry experience
- 4+ years
- Languages
- English, Portuguese
- Typical response
- Within a few hours
- Preferred engagement
- Small (individual or single team)
- Travel
- No travel
- Available in
- United States, Brazil
02
Proven on the network
Implementation records and publicly available work, where published.
0 AI Deployments delivered. No publicly identified implementations are available to inspect.
03
What Guidevit builds
Implementation domains listed on this profile. Evidence-supported items are identified separately below.
Specialisations
AI Agents
AI Evaluation & Testing
Autonomous Agents
Customer Support AI
RAG & Knowledge Retrieval
04
AI Deployments
Real AI implementations demonstrated through evidence.
No publicly identified Deployments are available to inspect.
05
Implementation capabilities
Evidence-supported means linked to a published Deployment or Case Study; other capabilities are Builder declared.
Evidence-supported
No published evidence linked yet.
Builder declared
- AI Agents
- AI Evaluation & Testing
- Autonomous Agents
- Customer Support AI
- RAG & Knowledge Retrieval
06
Technology
Technologies listed by the Builder, distinguished by published evidence.
Evidence-supported
No published evidence linked yet.
Builder declared
- Braintrust
- Claude Agent SDK
- FastAPI
- LangChain
- LangGraph
- LlamaIndex
- Model Context Protocol
- Neo4j
- PostgreSQL
- Python
- Qdrant
- pgvector
07
Industry experience
Published Deployments support the listed counts. Other industries are Builder declared.
Automotive
Builder declared
E-Commerce
Builder declared
Legal
Builder declared
Retail
Builder declared
Software & SaaS
Builder declared
08
Trust & verification
A closer look at how this profile and its evidence are reviewed.
Community Member
Published evidence is reviewed against the AI Deploy Network Evidence Standard.
How verification works
Verified Builder
Independently reviewed by AI Deploy Labs against the platform Evidence Standard.
How this is verifiedWhat the platform has established, and what remains builder-declared.
- Identity and organisation
- Identity checks confirm that the person presenting this profile controls the account and the professional identifiers attached to it, such as a LinkedIn profile. Organisation membership is confirmed by the organisation, not self-declared by the builder.
- AI Deployments
- An approved AI Deployment is an implementation record the builder submitted and the platform reviewed against the Evidence Standard before publication. Approval confirms the record is complete, coherent and consistent with the supporting material provided — it is not an audit of the client's systems.
- Business outcomes
- Outcome figures are recorded by the builder on the deployment or case study that produced them, and are reviewed at publication. Aggregates shown on this profile are calculated from those approved records only.
- Evidence
- Supporting evidence comes from the records the builder published on this platform: approved AI Deployments, published AI Case Studies, verification outcomes and organisation participation.
- Currency
- Older evidence continues to count and is never removed. Recency is presented separately through Professional Currency so that current expertise can be assessed alongside historical work.
- Traceability
- Every figure on this profile can be followed to the record that produced it. Open the linked AI Deployment or AI Case Study to read the outcome exactly as it was reported.
The platform does not independently audit client systems or financial records. How this is verified
Professional record and further evidence
Client Portfolio
Diverse industry experience
Nicio
Machine Learning • Legal
Brazil2026Co-founder of Nicio (nicio.com.br), an AI product for trademark law firms in Brazil. Each week, INPI publishes ~10,000 new marks; Nicio compares them with a firm's portfolio and sends only likely conflicts to a review queue, ordered by days left to oppose. - Built the platform end to end: weekly INPI ingestion, a proprietary spelling and sound model, and an AI step that decides what reaches the queue - Every alert explains its signals for and against - ~9.7M comparisons per edition for a 1,000-mark portfolio - Daily releases via my custom AI harness; I own the release decision
DealsmartAI
Agentic AI • Automotive
United States2025I build autonomous sales agents for automotive dealerships (dealsmartai.com). - Core agent workflows on Claude Agent SDK and FastAPI, with Neo4j graph memory for multi-turn context - Braintrust eval pipeline gating CI/CD: ≥95% accuracy and 100% on critical P0 cases - End-to-end tracing for production debugging - Compliance layers that trigger agent self-correction before delivery - Internal agent harness for planning, scaffolding, and adversarial verification Stack: Python, TypeScript, Claude Agent SDK, MCP, FastAPI, Braintrust, Neo4j
Netair
Automation • E-Commerce
Brazil2025I built an end-to-end AI data enrichment system for an automotive parts business, processing 60,000+ supplier products. Highlights: Data pipeline (acquisition, cleansing, normalization, enrichment) on Supabase (PostgreSQL), with Deep Research via Google ADK and n8n workflows Sales chatbot that cross-references part compatibility, combining web search with Neo4j graph memory for instant availability answers Real-time inventory sync across physical stock, ERP, and marketplaces, cutting stockout sales errors by 90%+ Stack: Python, n8n, Supabase, Neo4j, Google ADK
Rumo Logistics
Conversational AI • Logistics & Transportation
Brazil2024Enterprise AI assistant for 5,000+ users: 70% resolution rate, 86% fewer support escalations RAG architectures integrated with Jira, Active Directory, SAP Concur, and messaging platforms Real-time meeting analysis that validates engineering discussions against project docs 15,000+ monthly interactions: 89% fewer support calls, 88% fewer tickets
Zinv AI
Agentic AI • Legal
Brazil2023Responsible for innovation in legal process automation at a startup, involving tasks ranging from document retrieval to generating insights for new legal documents. I work on data collection and preparation for fine-tuning language models, and train these models on platforms such as AWS and Hugging Face. I develop RAG pipelines and implement indexing strategies, utilizing NoSQL and Vector Databases. I oversee the entire cycle of AI model customization and evaluation, aiming to deliver effective solutions aligned with business objectives.
Ready to discuss your AI implementation?
Tell Guidevit what you are trying to automate and the outcome you need. Your message goes directly to the builder, with a copy kept in your Organisation Workspace.