Carlos Corgozinho | Founder @ Dulyx

Privacy by designChanges need approvalWhere the company already isLess data, lower costAgents with toolsHarness with a human in the loopRAG and hybrid retrievalContext reductionReady-made integrationsEvals and observabilityPrivacy by designChanges need approvalWhere the company already isLess data, lower costAgents with toolsHarness with a human in the loopRAG and hybrid retrievalContext reductionReady-made integrationsEvals and observability

0+: years I've been building for the web
0+: integrations I built at Dulyx
0.0M+: lines of code I shipped

0: startup I founded, Dulyx
0+: systems I put into production in retail
0: e-commerces I integrated with an ERP
0: libraries I published on npm

Founder

I founded Dulyx, a layer between AI and your company's systems.

I built it so the model asks in natural language, Dulyx queries the ERP, CRM or API at the source and returns only the answer. Credentials never reach the model.

dulyx.com.br →

~/dulyx

Privacy by design

I keep credentials isolated inside the platform. The model only ever sees the question and the answer, and I log every operation.

Changes need approval

I let queries run straight through. Any change to an order, product or stock level I turn into a proposal that a human has to confirm.

Where the company already is

I put Dulyx on the company website and in any MCP-compatible assistant, with agents per area: sales, finance, support and logistics.

Less data, lower cost

I reduce the records before they reach the model and I bill per conversation, with no per-seat pricing.

fig. 01 · four principlesdulyx.com.br

AI Engineering

Models are the easy part. I build everything around them.

At Dulyx I build the infrastructure that makes an agent trustworthy inside a company: well-defined tools, data reduced before the model, human approval for every change and a full trace of each operation.

01Questioncompany site or an assistant over MCP
02Agentsales, finance, support or logistics
03Toolstyped queries and actions, one per intent
04SystemsERP, CRM and APIs at the source, credentials isolated
05Reductiononly what matters goes back to the model, with metadata
06Answeror a proposal waiting for approval

fig. 02 · how a Dulyx conversation flowssix steps

Agents with tools

I scope agents per area, each seeing only the tools of its domain. I give every tool a clear input and output contract, so the agent never invents data: it queries for it.

tool usemulti-agentMCP

Harness with a human in the loop

I built the runtime that routes calls, applies guardrails and turns every change into a proposal. Nothing moves in the ERP without a person confirming, and I record all of it.

human-in-the-loopaudit trailguardrails

RAG and hybrid retrieval

I do retrieval over documents and catalogs combining lexical, phonetic and embedding search. I created CAISS out of this: precision on product names, codes and abbreviations, exactly where vectors alone fall apart.

RAGembeddings@ccorgz/caiss

Context reduction

I trim and summarise records before they reach the model, with metadata that flags when a result set came back truncated. Fewer tokens, lower cost, less hallucination.

contextcostmetadata

Ready-made integrations

I built twenty-plus connectors for enterprise systems and channels, and I expose them as MCP servers for Claude, ChatGPT and other compatible assistants.

MCP serverERPCRM

Evals and observability

I trace every conversation: tools called, data returned, decisions taken. That is how I measure quality, compare models and improve prompts with evidence.

evalstracingquality

Canada · 2026

I spent four weeks in Canada, where Dulyx as a business was born.

I studied in Vancouver for three weeks, meeting people from all over the world and seeing first-hand what is being built outside Brazil. Then I spent one more week in Toronto and Guelph, living the local culture.

Event

Web Summit Vancouver 2026

I attended Web Summit: days exchanging experience with founders, engineers and investors from many countries, and watching Brazilians present their businesses on a global stage. That is where I had the idea of Dulyx as a business.

AttendeeStartupsNetworkingApplied AI

Certification

English C2.2 · EF Vancouver

I took a full-time in-person course at EF Education First and completed it at level C2.2, the highest on the scale. I had never taken an English course in Brazil: I learned from series, music, technology and work, then validated it abroad.

C2.2April–May 2026Full-time

“Technology, in the end, is about people, ideas and the connections we create along the way.”

Web Summit and three weeks in Vancouver Open on LinkedIn ↗

C2.2 certificate at EF Education First Open on LinkedIn ↗

fig. 03 · posted on LinkedInvancouver · 2026

Open source

I publish open source libraries on npm.

Each one came out of a real problem I hit in a real project.

packagewhat it doesinstall

reactivus

UI · React

My reusable React components (inputs, toasts, dialogs, tables) with a simple API. I use them to build screens fast without giving up consistency.

npm i reactivus

npm ↗Demo →

@ccorgz/caiss

Search · Similarity

My in-memory similarity search engine, with phonetics and inverted indexes. I built it to match product names full of abbreviations and typos, no embeddings required.

npm i @ccorgz/caiss

npm ↗

hmacpass

Security · Node

I wrote it to hash and verify passwords with HMAC-SHA512 and salt: lightweight, dependency-free and straight to the point.

npm i hmacpass

npm ↗GitHub ↗

typix

Validation · TypeScript

I wrote it to validate object types and values with custom rules, pairing strict type checking with content validation.

npm i typix

npm ↗GitHub ↗

fig. 04 · packagesnpmjs.com/~ccorgz

Work

I ship ERP-integrated commerce, end to end.

I built a full ecosystem: B2C store, B2B portal, backoffice and an operations app, all speaking to the same Oracle ERP.

B2C e-commerce

Minas Rural↗

I built the online store integrated with the company ERP: catalog, pricing, stock and orders in sync. Banners and product photos are editable in real time through Painel Web.

Next.jsAdonis.jsNode.jsOracleLinuxGTM