Portfolio · 2026

Lorena Kapitzky

I work the Forward Deployed Engineer way: I get inside the operation, understand the business problem and ship a working system. Finance, data and AI agents.

Python · FastAPINext.js · TypeScriptSQL · MySQLAI agentsPower BIAudit & controls
9screens in the dashboard
118automated tests
24pull requests shipped
5AI agents on the team
Main project

ERP-integrated management dashboard

Internal web app for an industrial distributor. It reads the company ERP in read-only mode and brings together what used to be scattered across reports.

  • Screens for customers, suppliers, inventory, sales pipeline, accounts payable and receivable.
  • Date-range filters, export and batch payment reminders.
  • Role-based login, audit trail and SELECT-only ERP access.
  • Tests and CI on every pull request.
FastAPINext.js 16MySQLSQLiteAlembicGitHub Actions
Illustration with sample data. The real dashboard runs on a private network.
Applied AI

An AI agent team that builds the dashboard

I set up a team of 5 agents that works from a Kanban board, talks to me on Slack and ships code through pull requests, only when tests pass.

  • The coordinator, Registro, splits requests into cards of up to 1 hour and assigns them.
  • Review is separate from merge, and related PRs merge as a group to avoid conflicts.
  • Scheduled jobs: morning summary, product questions and a weekly security scan.
  • AI model chosen per role to keep costs down.
  • Every agent mistake became a written rule the whole team follows.
Hermes AgentKanbanSlackGitHubSecond brain
Lorena asks on Slack Registro coordinator · Kanban PM writes requirements Builder codes and tests Reviewer reviews and approves Security weekly audit GitHub PR + CI Merge only when tests pass
How one of my requests becomes shipped code.
Day to day

Managing the team on Slack

I ask for features in plain language. Registro confirms the scope, opens the cards and tells me when the PR is ready for my approval.

Registro has its own pixel-art avatar so the coordinator stands out on Slack.

Sample conversation.
Infrastructure and security

Self-hosted, nothing exposed to the internet

I deployed the dashboard and the agents on a VPS that only my family can reach, over a private network.

  • Docker Compose, Tailscale with HTTPS, firewall and key-only SSH.
  • Secure bridge to the office ERP, open only to the server's private IP.
  • Read-only database user and a least-privilege GitHub token.
  • Snapshots, weekly backups and leak audits.
DockerTailscaleufwLinuxosv-scanner
VPS · FIREWALL Web dashboard (Next.js) API (FastAPI) + SQLite AI agent team Family phone and laptop ERP (MySQL) at the office Tailscale read-only No ports open to the internet. Private network access only.
Simplified architecture.
Data

ERP data analysis in SQL

I dug into the ERP database so every number on the dashboard matches the business.

  • Mapped Brazilian fiscal codes (CFOP) and operation types to separate sales from other outflows.
  • Validated counts and totals in the database and analyzed query plans.
  • Handled legacy data, such as zero dates, without changing the ERP.
MySQLDBeaverEXPLAIN
consulta.sql
SELECT n.natureza, COUNT(*) AS notas
FROM notas_saida s
JOIN naturezas n ON n.codigo = s.natureza
WHERE s.cfop IN ('5102', '6102')  -- sales only
GROUP BY n.natureza;
Simplified example. Table names are illustrative.
Automation

Public tender radar

A daily job that searches Brazil's national procurement portal (PNCP), filters for the products the company sells and posts a summary on Slack at 7 a.m.

PNCPSchedulingSlack
Sample summary.
Audit automation · Grant Thornton

From 180 PDF pages to a dashboard in 20 minutes

Every month, 3 PDFs of about 60 pages each arrived with the data for an outsourced-services contract. Extracting it by hand took 2 days.

  • Built a Python script that extracts the data automatically: 2 days became 20 minutes.
  • Sole owner of the Power BI dashboard used in the monthly stakeholder meeting: contract overruns in value or hours, and hours worked above the permitted limit.
  • Audited 30+ expense receipts and service contracts every month.
  • At a construction company, compared the payroll against the collective bargaining agreement and found site crews' overtime paid at a 100% premium when the agreement set 50%. Even the finance manager believed 100% was right; the premium was corrected.
PythonPDF extractionPower BIInternal audit
3 PDFs ~60 pages each Python automatic extraction Power BI dashboard Meeting stakeholders 2 days → 20 min time to extract the monthly data
Monthly flow, simplified. No client data shown.
How I work

Rules where it must be exact

Fixed rules for what has to be right every time, like what counts as a sale or how a payment matches an invoice. AI only where judgment helps.

Two checks on every AI output

First the format: required fields present and well formed. Then the business rules: the numbers have to make sense. Anything that fails goes to a person.

The system prepares, people approve

Automations draft, match and flag. A person reviews and decides, and the history stays recorded.

Measure the whole job

I start with a small proof of concept on real cases and measure accuracy and the review work left for the team, not just the model's hit rate.

Development

Python, FastAPI, TypeScript, Next.js, React, REST APIs, automated testing, Git and CI.

Data

SQL, MySQL, SQLite, ERP integration, data modeling and analysis, dashboards.

AI

Agent orchestration, prompt engineering, cost-aware model selection, reviewing AI-generated code.

Infrastructure and security

Docker, Linux, Tailscale, firewalls, least privilege, backups and leak audits.

Finance and audit

Internal audit, internal controls, risk assessment, credit risk, accounts receivable, public procurement.

Learning

Agentic AI and AI Python for Beginners (DeepLearning.AI), Forward Deployed Engineer (Udemy, in progress).

Open to work

Forward Deployed, Solutions and Implementation roles

In AI, fintech and B2B software. Remote or in the EU: I'm an Italian citizen, no visa needed.

All screens and messages on this page use sample data. The real systems run on a private network and hold confidential company data.