Skip to content
Kaique Savi

Kaique Savi / Head of Technology

AI in production,not in slides.

// the LLM suggests. the tests decide. production judges.

Full-stack engineer with 3+ years building production SaaS. Today I am Head of Technology at DELIPE and own the whole platform: TypeScript back ends (NestJS, Fastify, PostgreSQL), applied AI with Claude and OpenAI, MCP servers, and integrations with ERPs, WhatsApp and phone systems.

3+ yrsbuilding and running production SaaS→ experience
2×promoted in 3 years, from junior to Head of Technology→ career
5products in production that I built or led→ projects
FIAPstudying Artificial Intelligence; winner of Global Solution 2025→ education

01 / experience

From junior to head, same company.

I started out shipping landing pages; today I own the whole platform. Three years of real production, with customers using it the next morning.

Marketing, sales and technology products for internet service providers in Brazil and Latin America.

Jun 2026 — now

Campinas, Brazil

Head of Technology

Own architecture, infrastructure and delivery for the product suite: AI monitoring, project management, CRM and CMS.

  • Designed and built DELIPE - OS, the company’s operations platform: boards, automations, calendar and timesheets, contracts, campaigns, and integrations with WhatsApp, PBX and SSO.
  • Opened the platform to AI agents through an in-house MCP server with 109 read and write tools.
  • Ran a cloud cost audit and a phased plan that removed idle capacity and cut the monthly bill.
  • Rolled out error monitoring with AI triage that turns production errors into pull requests or issues.
  • Set the engineering standards: bounded-context modules, multi-tenant isolation, signed webhooks, rate limiting and credential rotation.
83modules in DELIPE - OS
109MCP tools
900+merged PRs

TypeScript · Fastify · PostgreSQL · Drizzle · Redis · BullMQ · React · Docker · Claude API

“If no test can fail, nothing is guaranteed to work.”

May 2025 — Jun 2026

Campinas, Brazil

Full-Stack Developer (mid-level)

The AI and integrations phase: products that talk to the customer’s systems and turn conversations into data.

  • Built the backend of the AI monitoring product: a multi-tenant platform that analyzes meetings, calls and support chats with Claude and produces coaching and KPIs for ISP teams.
  • Shipped an API that generates marketing creatives with AI, with queue workers and S3 storage.
  • Integrated ISP ERPs (IXC, Hubsoft, MK, SGP), support platforms (OPA Suite, Chat Mix), PBX telephony and the WhatsApp Cloud API.
  • Built ETL pipelines and n8n flows joining CRM, telephony and chat data, with Power BI dashboards.
317endpoints in AI monitoring
4ERPs integrated
~51MCP tools

NestJS · Drizzle · BullMQ · Python · LightGBM · n8n · Power BI · OpenAI · Whisper

“Every ERP has an API. Not every ERP has docs.”

May 2023 — Sep 2025

Sumaré, Brazil · hybrid

Full-Stack Developer (junior)

Where I learned to ship fast: brief in the morning, page live within days.

  • Shipped 150+ websites and landing pages for ISPs and for DELIPE itself, with Astro, React and TypeScript.
  • Built the front end of an NPS survey and a WooCommerce integration API.
  • Deployed and ran services with Docker, Docker Swarm and Linux.
150+websites and landing pages live

Astro · React · TypeScript · Docker Swarm · Linux

“A landing page is production too. The customer just sees it right away.”

Before code

Accounting and administrative assistant (2020–2023), alongside a technical degree in Business Administration. That is where I learned spreadsheets are the most used system in the world.

02 / projects

What runs in production, and what is just mine.

First the real work, which is private and runs every day. Then the after-hours projects, with open source code.

At DELIPE

Private code: no links, but the numbers come from git.

DELIPE - OS

platform

lead author

Where the company’s operations run all day; it replaced spreadsheets and Monday.com. A board engine with 21 column types (formula, mirror, rollup, relation), a visual automation engine with 18 step types, and an MCP server with 109 tools to run everything from Claude.

  • 83 modules
  • ~840 routes
  • 159 migrations
  • 900+ PRs
  • ~400k lines

Fastify · PostgreSQL · Drizzle · BullMQ · React 19 · React Flow · MCP

private

AI monitoring for ISPs

SaaS

led AI, MCP and churn prediction

Multi-tenant SaaS used by 200+ ISPs. The AI scores every support and sales conversation (quality, BANT, DISC, objections), predicts churn with LightGBM and answers managers about their ERP data through MCP, with a per-customer cost cap.

  • 52 modules
  • 317 endpoints
  • ~51 MCP tools
  • schema per tenant

NestJS · PostgreSQL · BullMQ · Claude · OpenAI · Python · LightGBM

private

AI monitoring

AI

lead author

The AI that monitors DELIPE’s sales team: it analyzes sales meetings and phone calls, joins them with targets and ad data, and raises alerts. A rep sends a WhatsApp voice note and the lead shows up in the CRM, ready.

  • 26 modules
  • 113 endpoints
  • 23 MCP tools

NestJS · Supabase · Meilisearch · Whisper · Claude · BullMQ

private

GEO · AI visibility

AI

module I built in the team’s Hub

Measures how ChatGPT, Claude, Gemini and Perplexity talk about an ISP’s brand, city by city: ranking against competitors, cited sources and factual errors, ending in an action plan.

  • 4 AI engines
  • 56 MCP tools
  • OAuth 2.1 + PKCE

Fastify · PostgreSQL (RLS) · Redis · MCP

private

Results Dashboard

data

Proves to the ISP that marketing worked: joins ad spend with sales and activations pulled straight from 4 ERPs and projects the ROI.

  • IXC, Hubsoft, MK and SGP
  • Google and Meta Ads
  • backups every 15 min

Express · Prisma · PostgreSQL · Playwright · Claude

private

Also: a LangGraph agent on top of Monday.com, a WhatsApp group auditor and an AI artwork generator.

After hours

Open source on GitHub.

01

isp-analytics-mcp

MCP

MCP server that gives an agent business-level tools over a fictional ISP (churn, delinquency, cohorts, sales) instead of a raw database dump.

Python · MCP · SQL

open
02

meeting-insights-ai

LLM

Agent that turns a meeting transcript into validated JSON (decisions, action items with owner and due date, risks), with an evaluation harness.

Python · LLM · evals

open
03

audio-transcriber

tool

Transcribes video and audio with timestamps using Whisper, splitting large files into chunks.

Node.js · ffmpeg · Whisper

open
see all repositories ↗

03 / education

Artificial Intelligence at FIAP.

An AI degree taken alongside my job. The program is project-based: every phase ships something that works, with data, a model and documentation, usually with agriculture as the setting.

Degree
Artificial Intelligence (undergraduate)
School
FIAP, São Paulo, Brazil
Period
Jan 2026 — Jan 2028 · in progress
Subjects
machine learning, data engineering, computer vision, IoT and cloud

award

Winner of FIAP Global Solution 2025 · Future at Work

I won with the fintech API. Global Solution is FIAP’s end-of-semester challenge; this edition’s theme was Future at Work. Certified on Nov 11, 2025.

see the winning project ↓

Course projects

Open source on GitHub.

01

Fintech API

Javawinner · Global Solution 2025

The project that won Global Solution 2025 (Future at Work): a REST API for a digital bank, with users, accounts, cards, transactions, investments and rewards, plus JWT authentication.

Java · Spring Boot · Spring Security · JWT

open
02

OrbiTerra

Global Solution

Water-stress risk for coffee crops fusing satellite data (NDVI/LST) with an ESP32 soil sensor. With the sensor, Random Forest accuracy goes from 83.9% to 97.4% under temporal validation.

Python · XGBoost · Random Forest · ESP32

open
03

AgroGuard IA

Enterprise Challenge

Sompo Insurance × FIAP challenge over 4 sprints: risk prediction for farm equipment with an API, database, dashboard, model card and 68 tests. I ran it solo from sprint 3 on.

FastAPI · PostgreSQL · Streamlit · Docker

open
04

Computer vision with YOLO

phase 6

YOLOv5 trained on a dataset I built, compared with pre-trained YOLO (COCO) and a CNN built from scratch.

PyTorch · YOLOv5 · CNN

open
05

Crop yield prediction

phase 5

Clustering, outlier detection and 5 regression models to predict crop yield, plus an AWS cost comparison between São Paulo and Northern Virginia.

Python · scikit-learn · AWS

open
06

Crop recommendation

phase 3

Exploratory analysis and predictive models over 2,200 soil and climate samples across 22 crops, as a team.

Python · pandas · scikit-learn

open
07

Smart irrigation with ESP32

phase 2

Reads soil moisture, pH and nutrients (NPK) and runs the water pump on its own. Circuit simulated on Wokwi.

ESP32 · C++ · PlatformIO · Wokwi

open

Before FIAP

Technical high school degree in Business Administration, Escola Municipal Dr. Leandro Franceschini (2021–2024).

04 / toolbox

What I reach for.

Grouped by what it does, not by this month’s hype.

Backend

  • TypeScript
  • Node.js
  • NestJS
  • Fastify
  • PostgreSQL
  • Drizzle
  • Prisma
  • Redis
  • BullMQ
  • Zod

Frontend

  • React
  • Next.js
  • Astro
  • Vite
  • Tailwind
  • TanStack Query

Data & AI

  • Python
  • Claude API
  • OpenAI
  • Whisper
  • MCP
  • LightGBM
  • XGBoost
  • SHAP
  • n8n
  • Power BI
  • BigQuery

Infra

  • Docker
  • Kamal
  • Cloudflare
  • Railway
  • Supabase
  • GlitchTip / Sentry
  • Linux

05 / AI skills

I also write for agents.

Skills are instruction folders that teach a coding agent (Claude Code) to do a job the right way, every time. I wrote these and use them every week.

Private repository. Here is what each one does.

  • /blindar

    Backend hardening across 6 axes (availability, anti-replay, credentials, authorization, audit and app identity). Diagnoses with file-and-line evidence and fixes by severity.

  • /ship

    End-to-end delivery with subagents: plan, code, local tests, review and validation in a real browser. House rule: no evidence, it didn’t happen.

  • /observabilidade

    Installs error capture on back end and front end, plus a pipeline where AI triages each error and opens a PR or issue on its own.

  • /raptor

    Runs security analysis (Semgrep, CodeQL and dependency checks) on any repository, inside Docker.

  • /rastreabilidade-uso

    Home-grown usage analytics, no SaaS: pages, actions and filters tied to the signed-in user, with an admin-only dashboard.

  • /tema-claro

    Adds light, dark and system themes to a React + Tailwind app, with no flash on load.

  • /transcrever-audio

    Local transcription with Whisper (free and offline), with automatic fallback to the API.

  • /video-narrado

    Generates a presentation video with animated slides and narration synced sentence by sentence.

06 / contact

Let’s talk.

Want to talk about a project, swap ideas on AI in production, or just learn more about my work? Reach out, I reply fast.

kaiquesavi1@gmail.com

kaique@portfolio: ~