Curitiba, Brazil —
AI Engineer · Founder of Dark Marlin · Curitiba
I'm João Moreti — computer engineer, software developer and founder of Dark Marlin. I build complete platforms — from architecture to deploy — with real users and real payments: web systems, integrations, data pipelines, and AI that is evaluated and kept under control.
There are currently 8.3 billion people in the world estimate from the UN World Population Prospects 2024 (medium variant)
Real systems, in production or active use.
A multi-agent system made of one director and several managers: the director delegates to the managers, and the managers delegate to their agents. A desktop app for the day-to-day of a small business, no terminal, no prompt writing. 4 clients using the service.
Turns scattered auction listings into a queryable base that keeps what the portals delete: 2,841 properties tracked across PR, RS and SC, 4,249 price snapshots and 2,788 auction outcomes preserved. Automated weekly collection, a data-quality layer that quarantines incomplete records before they become a number, and panels for coverage, freshness and price movement.
Platform for an in-person business network, live with real payments since May 2026: ticketing, membership club, city groups and per-franchise finance, with tenants isolated by PostgreSQL Row-Level Security. Architecture and half the code, as a pair.
Quoting and production ERP for a joinery shop: it replaces rule-of-thumb markup with industrial costing, versions every proposal and turns approval into instalments in the finance module — with the calculation shown behind every price.
Three Model Context Protocol servers, in Python, that give an assistant real tools: academic search across 25 sources with deduplication by DOI and arXiv id (15 tools, 275 tests), read access to the paperswithcode.co API (31 tools) and a Telegram bot that sends text and files (6 tools). Two of them open source.
Internal system of a third-party property manager: developments and housing units, owners, guests, occupancy calendar, cleanings with video evidence, a process engine the managers configure themselves, 129 granular permissions with an audit trail and an AI assistant bound by the same permissions. Ongoing consultancy since 2026.
Customer-service bot on the official WhatsApp Business API: a FastAPI webhook, LLM replies that carry the conversation history kept in Supabase, hexagonal architecture with the message core under test. Private code, December 2025 to January 2026.
From a topic to a published vertical video with nobody in the middle: LLM script, cloned-voice narration, imagery driven through the browser, charts drawn in code, karaoke captions and a lip-synced mascot. About 30 minutes per video on a desktop GPU; 38 videos published.
Single-product store for the Octus-Naut, a collectible toyart in a limited run of 100 numbered pieces, hand-modelled and hand-painted. The piece and its craft process, plus a purchase flow designed end to end — delivery, payment and review — on desktop and mobile. Freelance project.
Developer and consultant, from first internship to founder.
Internship followed by 5 months as a software developer. Machine learning projects for crop yield forecasting: data collection and cleaning, model evaluation and data analysis; ETL pipelines, REST APIs and test-driven practices.
Features on an ASP.NET MVC system; query and ETL optimization.
Freelance projects: RAG chatbots with vector databases, LLM integrations in production, dashboards and e-commerce landing pages.
Ongoing consulting for a hospitality company: internal management platform and AI initiatives, from architecture to cloud deployment.
Software studio behind Emais Network — a SaaS live in production with real payments — and other client platforms.
A computer engineer focused on solving problems — with AI where it earns its place, and all the way to production.
Hands-on AI engineering. Evaluating models: an LLM-as-judge scoring another model's output, golden sets with recall and precision, an OCR benchmark measured by CER/WER, and live contract tests against real APIs. Controlling their output: structured JSON outputs, validation with Pydantic and Zod, deterministic anti-hallucination guardrails, retries that hand the error back to the model, prompt engineering, and tool use / function calling with permissions checked before every tool. I build agents with tool loops and MCP servers.
On the data side: ingestion and ETL pipelines with scraping, an immutable raw layer (sha256), schema validation with quarantine instead of silent drops, key-based deduplication, idempotency, scheduled batch jobs, data modelling and migrations in PostgreSQL — with Row-Level Security to isolate tenants.
Founder of Dark Marlin, the software studio under which I deliver complete platforms for clients.
