Salvador Ferreira

Deloitte

AI & Data Analyst · Technology & Transformation (Financial Services) · Sep 2026 – Present · Amoreiras

Building an AI agent that migrates legacy data warehouses to dbt, in a team where multi-agent workflows made delivery about 10x faster.

Python · SQL · dbt · T-SQL · SSIS · LLM agents

• Optimised multi-agent AI workflows, accelerating the team’s rate of delivery by ~10x.
• Built SQL and Python data transformations in dbt to support database migrations.

AI agent for migrating legacy data warehouses

I am developing an AI agent that automates the migration of legacy data warehouses, built on T-SQL stored procedures and SSIS packages, to a modern dbt architecture. The agent works as a multi-stage pipeline. First, it cleans the SQL code of comments and noise while preserving the business logic. Then it classifies each procedure by migration complexity and analyses its dependencies, execution paths and data transformations in depth. From that analysis, it generates synthetic test data without ever accessing real data. Finally, it produces and validates the dbt models in a medallion architecture and automatically fixes the errors it finds. To ensure security and traceability, it includes guardrails that control the commands it runs and record every action in an audit log. The goal is to drastically reduce the manual effort of migrations that typically involve hundreds of procedures, without losing rigour or quality.

Correntes de Memória

AI Engineer & Financial Controller · Jun 2025 – Present · Areeiro

Replacement orders and payments agent

An LLM agent that reads caregiver-replacement orders (PDF), updates the care-management system and reconciles the payments.

Python · LLMs

Developing an LLM-powered agent that reads unstructured caregiver-replacement orders (PDF), updates the internal care-management system and reconciles caregiver payments automatically, replacing an end-to-end manual process.

Every caregiver-replacement order arrives as an unstructured PDF. Until now each one went through the same manual process from end to end: someone read it, entered the change in the internal care-management system, and then reconciled the caregiver’s payments by hand.

The agent takes that over. It reads the order with an LLM, updates the care-management system and reconciles the payments on its own. I am building it in Python; it is one of the 10+ AI agents and automations I have built at Correntes de Memória, which together cut the team’s manual workload by about 6x.

Automations and financial operations

10+ AI agents and automations that cut the team's manual workload about 6x, alongside the day-to-day finance operations.

Python · LLMs

• Built 10+ AI agents and automations (Python, LLMs) for repetitive operational tasks, cutting manual workload ~6x and freeing the team to focus on client retention and recurring revenue.
• Automating the shared service-quality tracking spreadsheet used by the care team, so quality data is captured consistently instead of by hand.
• Own day-to-day finance operations: 1,000+ invoices, receipts and payment confirmations recorded, verified and reconciled across accounts payable and receivable.

Happy Code Portugal

Programming Instructor · Jan 2025 – May 2025 · Parque das Nações

Taught Python, SQL and Excel to learners aged 3 to 17.

Python · SQL · Excel

Delivered Python, SQL and Excel training to learners aged 3–17, adapting technical content to very different levels of experience.

Iscte Junior Consulting

Analytics Consultant & Vice-President of the Fiscal Council · Oct 2023 – Jan 2025 · Cidade Universitária

50+ databases, dashboards and reports that turned raw data into recommendations for clients and internal teams.

Power BI · Excel

• Built 50+ databases, dashboards and reports (Power BI, Excel) for client projects and internal teams, turning raw data into clear recommendations for decision-makers.
• Contributed to the organisation’s ISO 9001 certification (quality management and process standardisation) alongside the analytics department.

NOVA IMS – Information Management School

MSc in Risk Management · 2026 – 2028 (expected)

Master's in Risk Management, including machine learning and AI applied to financial risk.

Master’s programme on understanding, modelling and managing risk in complex settings, with a focus on the financial sector: credit, market and liquidity risk, financial derivatives, actuarial science, banking and insurance regulation, and machine learning and AI applied to risk management.

ISCTE – University Institute of Lisbon

BSc in Data Science · 2023 – 2026

Bachelor's in Data Science: programming, statistics, databases, machine learning and deep learning.

Three-year bachelor’s degree (180 ECTS) run by Iscte’s Business School and School of Technology and Architecture. It covers programming in Python, mathematics and statistics, databases and big data, supervised and unsupervised machine learning, deep learning, and data security, ethics and privacy, and ends with an applied project proposed by an outside organisation.

This website

Personal project

This map of Lisbon: 79,000 buildings from open data drawn in the browser, with missing heights predicted by a model.

Astro · React · Three.js · React Three Fiber · TypeScript · Python · GeoPandas · scikit-learn

A map of Lisbon you can fly over, with the places I have worked and studied lit up in it. Everything you see is drawn in the browser from open data: there are no photos or map tiles.

• The city. 79,000 building footprints from OpenStreetMap, raised to their height (from the number of floors where it is mapped). A Python pipeline (GeoPandas, Shapely) cleans them, cuts them to the area and packs the whole city, streets and parks into under 4 MB of compact binary files. Trees and green spaces come from the Câmara Municipal de Lisboa’s open data.
• Heights, learnt. About 70% of the buildings have no height or number of floors in OpenStreetMap. A gradient-boosted model (scikit-learn), trained on the 23,819 that do, predicts the rest from each building’s shape and kind and from its surroundings, rounded to whole floors. Tested on parts of the city it had not seen (spatial cross-validation, 600 m squares held out), it is off by 2.3 m on average, against 2.8 m for the rule it replaced (the median of the nearest mapped buildings), and puts 84% of buildings within one floor of their real height, against 78%. It helps most with low buildings; above 40 m both are about 20 m off (chart below).
• Drawing it. Astro serves a static page; the city is a React island using Three.js through React Three Fiber. Roofs, windows and shop fronts are made by shaders, not textures, and anything that repeats (trees, cars, lamps, people) is drawn as one instanced mesh.
• A city that lives. Cars follow the real streets and stop at the 2,294 traffic lights and 8,354 zebra crossings mapped in OpenStreetMap; buses run on 80 real bus lines; trains, trams and funiculars on their tracks.
• Light and weather. The sun’s position over Lisbon is worked out for the current minute, so the city follows the real day and night, and the weather is the real weather, from Open-Meteo.
• Running smoothly. Detail is drawn only near the camera, and the page lowers its own quality if a computer struggles to keep up.

Designed by me and built with Claude Code, an AI coding agent.

Source code

LearnVestor

Personal Project - Financial Literacy App

A free app that teaches investing through a virtual portfolio and lessons.

Flutter (Dart) · PostgreSQL · Supabase · JavaScript

Learn investing through gamification with LearnVestor, a free app with no account needed. Track your virtual portfolio, complete lessons, and build financial literacy.

WebsiteTry the demoApp repositoryWebsite repository