Curriculum Vitae

I am a Principal Data & Applied Scientist at Microsoft focused on large language models, search quality, prompt engineering, agentic systems, and AI-assisted programming. I have worked continuously with LLMs since the GPT-3 era, before ChatGPT existed, and have built numerous tools, prototypes, product features, and evaluation methods with the technology. I am also the creator and lead of TinyTroupe, an open-source LLM-powered multiagent persona simulation project with 7,000+ GitHub stars.

I hold a PhD in Computer Science from the University of São Paulo and Université Paris-Sud. My broader experience spans peer-reviewed research, entrepreneurship, Data Science consulting, technical leadership, Natural Language Understanding, Information Retrieval, Machine Learning, formal methods, and behavioral simulation.

LinkedIn profile ↗ Current professional profile and network.

Education

2006 – 2012 | PhD in Computer Science

Universidade de São Paulo (Brazil) / Université Paris-Sud (France)

Thesis
Verification of behaviourist multi-agent systems by means of formally guided simulations

URLs
Full text (USP version) Full text (Paris-Sud version)

I had scholarships from both CNPq (in Brazil) and CAPES (in France), as well as a scholarship provided directly by Université Paris-Sud.

2001 – 2005 | Bachelor of Computer Science

Universidade de São Paulo (Brazil)

1998 – 2000 | High school

Colégio Bandeirantes

Professional experience

Sep 2025 – present | Microsoft — Copilots & Agents Core

Principal Data & Applied Scientist
São Paulo, Brazil · Remote

I am part of Copilots & Agents Core (CACore), developing M365 Copilot Chat. My primary responsibility is improving search quality, using LLM-based techniques throughout development and evaluation.

• Apply prompt engineering, agentic approaches, and other LLM techniques to M365 Copilot Chat.
• Develop and evaluate methods for improving search, retrieval, and response quality.
• Use AI-assisted programming extensively to build and iterate on product features, tools, and prototypes, including long-running Claude Code and Copilot CLI sessions that often continue overnight.

Nov 2020 – Sep 2025 | Microsoft — Search, Assistant and Intelligence

Sr. Data & Applied Scientist
São Paulo, Brazil

I worked primarily in the Search Relevance team within MSAI, developing the enterprise version of Copilot—now M365 Copilot. Before that, I worked in the User Understanding team, whose Natural Language Understanding work affected products used by millions of people.

My responsibilities and accomplishments included:

• Prompt engineering for product features and auxiliary tools, including the critical capability that combines Web and Work search results.
• Opportunity analyses, data exploration, and experimentation with the quality of Copilot responses.
• Creation and leadership of TinyTroupe, an open-source LLM-powered multiagent persona simulator that now has 7,000+ GitHub stars.
• Extensive work with GPT-3, GPT-4, and related OpenAI technologies across prototypes, tools, and studies in NLU, Information Retrieval, analytics, Data Science, and other domains.
• Winning the global internal CTO Codex Innovation Challenge 2021/2022 with Tabular Data Processing with Natural Language; the challenge had an 8% win rate.
• Writing an in-depth LLM application development tutorial in early 2023.
• Building NLU solutions for intent classification, information and entity extraction, query understanding, and query rewriting.
• Applying Information Retrieval techniques and metrics, as well as analyzing user behavior through multiple methods.
• Serving as one of the leaders of the Brazil instance of Microsoft's Global Hackathon 2023.

2019 – 2020 | Funcional Health Tech (Health Analytics business unit)

Data Science Manager

https://www.funcionalhealthtech.com.br/

I was responsible for managing the Data Science area of the Health Analytics business unit. This included leading a team of three Data Scientists and one Solutions Architect, developing strategic technology and methods, and promoting good practices across the company and its customers. Important initiatives I led included:

• Established an agile (Scrum-based) work routine for the team;
• Designed and delivered workshops for internal business customers, in order to elicit the most valuable and feasible Data Science use cases;
• Designed and established a Data Science process for systematic and gradual work. This can be used both in consulting engagements and in product development;
• Compiled the most relevant Data Science use cases into more abstract and reusable scenarios, which allows more efficient use of the team's time;
• Infused good Software Engineering practices in Data Science work, in order to make it more reusable and automated;
• Promoted marketing work, which resulted in a favorable press release in top newspapers and magazines;
• Directly provided Data Science advisory as needed by other areas of the company.

2016 – 2019 | Dell Technologies (Consulting Services)

Data Scientist

https://www.dellemc.com/en-us/services/professional-services/bigdata.htm

As a Data Science consultant, I designed and implemented machine learning models for corporate customers, in particular through the analysis of large-scale, distributed, datasets (i.e., "big data"), typically using a Python stack. General expertise exercised involved:

• Hadoop-based tools (mainly Spark and Hive).
• machine learning libraries (e.g., scikit-learn, keras) and notebook environments (i.e., Jupyter and Zeppelin).
• worked directly with customers in order to assess relevant business problems and propose solutions.
• worked mostly in an agile manner (Scrum).

More concretely, I worked mainly in problems for the financial sector, such as:

• customer analysis through various techniques, including clustering and behavior forecast.
• cash flow forecasts using Recurrent Neural Networks (RNNs) and Decision Trees.
• recommendation system for a major Brazilian bank through (a simple version of) Reinforcement Learning, in collaboration with various other Data Scientists, Data Engineers and Solutions Architects.
• credit score analysis using non-traditional methods, in collaboration with other Data Scientists.

Other engagements included:

• technical advice with respect to the Data Science practices of a major European telecommunications operator.
• classification of crops using satellite imagery and Deep Learning techniques for a biotechnology company.
• advice for a major oil company regarding possible Data Science applications involving seismic data.

Further responsibilities:

• mentoring of junior members of the Data Science team.
• technical guidance to development team, particularly with respect to software testing.
• helped in various project proposals, with a particular interest in developing the customers' internal Data Science capabilities.

Entrepreneurship

2015 – 2017 | Salem Sistemas

Founder

www.salem-sistemas.com

My software company, through which I provided consulting services and developed new technologies. My main technical achievement here was called Empirica, an experimental software designed to optimize human behavior through the analysis of online user interactions. This research and development is currently paused.

2012 – 2016 | Liberalis

Founder

Liberalis project record

A startup product that I conceived, implemented and marketed after my PhD. It is an online tool through which professionals (e.g., psychologists, lawyers, architects, etc.) can easily create their own websites. The generated websites adapt to quantity and type of content inserted, without any need of manual layout or design configuration. Liberalis operated under a "freemium" model: the basic offer is free (with ads), and a premium offer (with a number of advantages) is charged. This system had over 2500 users and was implemented in Ruby on Rails.

In the press

Profissional liberal ganha site pessoal de graça com a Liberalis

Empreendedor deve avaliar custo antes de abrir escritório

Metade dos microempreendedores usa a própria casa para fazer negócios

Note
The payment subsystem of Liberalis was the inspiration for a formal verification software that I developed as an independent library and is now open sourced (https://github.com/paulosalem/verum). I also produced a related academic publication (see below; published in May/2016).

2005 – 2009 | Distincards

Co-founder

www.distincards.com

An e-cards website for people interested in art, literature and philosophy. "Distinct cards for distinct people."

Languages

English

Fluent. TOEFL iBT (116/120). IELTS (8.5/9).

French

TCF B2.

Good oral and written understanding; good oral expression; reasonable written expression. I lived in Paris for more than a year during my doctorate.

Portuguese

Native.