Infinity Loop Start a project

Software built from a blank page, and built to run every day.

Infinity Loop builds products, internal systems and apps from the first line of code to daily operation. Current engineering, production AI where it earns its place, and evidence before anything is called done.

Start a project See production systems
Production evidence
Running systems
  1. 01

    TalkERP

    ERP approval

    WhatsApp → Web → Protheus
    In production
  2. 02

    Inhive

    Grounded answers

    PDFs → RAG → Field teams
    100+ users
  3. 03

    Woman's Academy

    One platform

    Web + iOS + Android
    5,000+ students

Three of ten delivered systems. Architecture and numbers shown here are drawn from production work.

In production

Three systems, drawn the way they actually work. The architecture makes each decision legible without a login.

TalkERP

Wholesale distribution, Totvs Protheus

Sales order approval that lives in WhatsApp, on top of an ERP that was never built for it.

A multi tenant agentic layer. Each tenant's ERP credential is encrypted in Firestore and wrapped in an RBAC permission envelope, so an approver only ever sees the orders their role allows. The agent service runs privately on Cloud Run behind ID token authentication, with secrets in Secret Manager and deploys through Cloud Build.

Tools return JSON directly with stop_after_tool_call. Letting the model rewrite the order listing cost around 600 output tokens per order, and it was rewriting data it had just been handed.

  • Fastify
  • TypeScript
  • React
  • Vite
  • Python
  • Agno AgentOS
  • Firestore
  • Cloud Run
  • Secret Manager
TalkERP approval flow across WhatsApp and the web A sales order request arrives in WhatsApp with a signed link. The approver opens a web page and selects orders there. The confirmation is sent back inside WhatsApp. The gateway then writes the approval to Totvs Protheus through the NoAlcada API. WHATSAPP DECISION 01 Approval request with a signed link 03 Confirm approval inside the chat, never on the page 04 Write to Protheus NoAlcada API, tenant credential WEB SELECTION 02 Select orders the agent lists them, filtered by RBAC role selection returns to the chat order data

Scroll sideways to see the whole flow.

The channel split is deliberate. A page can list two hundred orders; a chat cannot. A chat has the approver's identity and their habit; a page does not. So the listing lives on the web and the decision stays where the approver already is.

Inhive

Facility cleaning and maintenance

Every procedure a crew needs, answerable in plain language, with the source attached.

It is a sector with high turnover where the procedures exist only as PDFs, out of reach of the person standing at the post. Inhive centralizes standard procedures, norms, training and the rules of each contract, gives every organization its own knowledge base and specialized agents per area, and keeps access scoped by client and by team.

100+ users across two paying clients, in production
  • React
  • TypeScript
  • Tailwind
  • shadcn/ui
  • Supabase
  • pgvector
  • Stripe
  • n8n
  • Capacitor
  • GitHub Actions
Inhive ingestion and answer path Standard operating procedures, contract rules and training material are ingested through an n8n pipeline into a per-organization store on Supabase with pgvector. A worker asks a question in plain language and the specialized agent answers with the source document attached. INGESTION ONE STORE PER ORGANIZATION Standard procedures Contract rules Training material n8n pipeline chunk, embed, index Its own knowledge base Supabase, PostgreSQL, pgvector access scoped by client and by team ANSWER SOURCE ATTACHED Asked in plain language from the post, on a phone, in Portuguese, English or Spanish Agent for that area retrieval scoped to the organization and the contract Answer plus source the document it came from travels with every answer

Scroll sideways to see the whole flow.

In a sector with this much turnover, an answer without its source is a rumor. The document that produced the answer is part of the answer, so a supervisor can check it against the contract without leaving the thread.

Woman's Academy

Online education

The whole platform behind a course business, not a members area bolted onto a checkout.

Members area, an app of its own on iOS, Android and web, an admin panel, and payments through Eduzz, all on one Node.js backend. A student is one record, so access, content and reporting cannot drift apart between surfaces.

5,000+ students on the platform
  • Node.js
  • iOS
  • Android
  • Web
  • Eduzz
Woman's Academy platform surfaces Enrollment and payment arrive from Eduzz into a single Node.js backend, which serves the members area on the web, the app on iOS and Android, and the admin panel. PLATFORM Eduzz payment and enrollment One Node.js backend enrollment and access course content reporting Members area on the web iOS app its own, not a wrapper Android app same platform behind it Admin panel students, content, sales ONE PLACE WHERE A STUDENT EXISTS
Course businesses usually run on four disconnected tools and reconcile them by hand. Here the student exists once. Enrollment, access, content and reporting are the same records, so an app and the admin panel cannot disagree about who paid for what.

Also shipped

  1. TronST

    Customer service

    Chatbots and AI agents answering support on WhatsApp.

    Running in production on the official WhatsApp Business API, with the conversation logic orchestrated in n8n and answers generated through LLM APIs.

    • n8n
    • LLM APIs
    • WhatsApp Business API

  2. Merchand

    Supermarket merchandising

    Scheduling, field tracking and performance for merchandising teams.

    Visits are planned in the panel, checked in and out from the aisle, and reported back the same day. The WhatsApp API carries the parts of the loop that never survive a separate app.

    • Firebase
    • Google Cloud
    • WhatsApp API

  3. Vital Sense

    Health

    A tongue photograph read by a multimodal model into a health report.

    Published on the App Store and Google Play and selling through the store wallets, StoreKit and Google Play Billing, with an admin panel behind it. This is multimodal API integration, not a model we trained.

    • Flutter
    • Multimodal LLM
    • StoreKit
    • Play Billing

  4. Fiabesa

    Raffia and packaging industry

    500+
    inspections per day, 24/7

    Quality inspection on the factory floor, written back into the ERP.

    Extrusion, weaving, lamination, bagging and RNC inspections captured on shift and written back into Totvs Protheus through Firebase, on a Node.js backend.

    • Node.js
    • Firebase
    • Totvs Protheus

  5. Agrobiota

    Agricultural laboratory

    Service orders connecting the soil sampling team to the lab.

    Field collection and laboratory intake kept on one order, integrated with Pipefy over its GraphQL API so the process stops losing steps between the two teams.

    • Firebase
    • Azure
    • Pipefy GraphQL

  6. Kaplen

    Card sales reconciliation, cofounded and built

    200+
    clients reconciled

    Reconciliation of credit card sales for retailers and airlines.

    Microservices on AWS with Java and Spring: Elastic Beanstalk, SQS, SNS, RDS, Lambda, API Gateway, CloudFront and DynamoDB, with Redis on the hot path and AngularJS on the front. Sized to reconcile the card sales of more than 200 clients.

    • AWS
    • Java
    • Spring
    • Lambda
    • SQS
    • DynamoDB
    • Redis
    • AngularJS

  7. Segoo.br

    Insurance brokerage software, cofounded and built

    500+
    daily users

    Quotes, policies and renewals for insurance brokers.

    Microservices talking to each other asynchronously, so a slow insurer API never blocks a broker mid quote. Sized for more than 50 brokerages.

    • Microservices
    • Async messaging

How we build

A task is not done when the code is written. It is done when there is evidence it meets the acceptance criteria.

Development here is agent oriented, and on TalkERP that is a written process rather than a habit.

Claude Code, Codex and Cursor are in the daily loop. Around them sits a team of agents with defined roles, a board they all write to, security conventions committed to the repository, and a rule for what counts as proof. Plenty of shops use AI to write code. Far fewer run a process with it.

  1. Lead
  2. UX
  3. Tech Lead
  4. Developer
  5. Reviewer
  6. QA

Founder

Diego Lima has been building software for more than fifteen years, most of it as the engineer accountable for the architecture. Twice a cofounder and CTO, twice from the first line of code.

15+ years shipping software people depend on at work
Kaplen Cofounder and CTO, acquired by Itaú Unibanco in 2015
What a card operator says it will pay, matched line by line against what the store actually sold and against what finally landed in the bank. Retailers and airlines run that reconciliation every day. Diego built the architecture and the core from scratch and led the engineering team that took the product to market.
Itaú Unibanco After the acquisition
The product kept running while it moved onto the bank's own infrastructure. Diego led that migration, designed the architecture template the bank's teams then built against, and formed the internal team that took the platform over. Same product, a different operation around it.
Segoo.br Cofounder and CTO
Quotes, policies and renewals for insurance brokers, in a market where the quote a broker gives a client depends on insurer systems answering at their own pace. Built so a slow one never holds the broker up in front of the client.

Tell us what needs to reach production.

diego.lima@infinityloop.app

Write in English or Portuguese. Describe the system, the constraint you are working against, and the date that matters. You will get an answer from the person who would build it.