Artur B. Carneiro

Artur B. Carneiro

AI researcher at Stanford. Working on efficient reasoning, mechanistic interpretability, and making large models smaller and faster.

Experience

Stanford SISL (SAIL)AI Research
Nov 2024–Aug 2025

Research in Mykel Kochenderfer's lab at the Stanford Intelligent Systems Laboratory (SISL).

Stanford University School of EngineeringTA & Section Leader
Jan 2023–Present

Currently TA for AA 228/CS 238: Decision Making Under Uncertainty. Previously Senior Section Leader for CS 106A, CS 106AX, and CS 106B.

Instagram Machine LearningSoftware Engineering Intern
Jun–Sep 2025

Improved the offline simulation pipeline for machine learning algorithms powering Instagram's Feed and Reels recommender systems, working on how models are evaluated at scale in production environments.

Meta PlatformsSoftware Engineering Intern
Jun–Aug 2024

Built and shipped UI components for the Facebook Design System, impacting 100M+ Marketplace users. Collaborated with one of Meta's leading accessibility experts to develop a novel solution to a longstanding accessibility problem, establishing a new standard across the Facebook app. Went deep enough into the stack to uncover bugs in underlying frameworks — all changes are now in production for all Facebook users.

PulsarSoftware Engineering Intern, Applied AI
Jun–Aug 2023

Built a natural-language interface for querying data, manuals, and datasheets across 2,000 industrial machines using Retrieval-Augmented Generation and LLMs. Engineered an API to orchestrate a conversational experience within a messaging system, allowing users to interact from their mobile devices.

Education JourneyCo-founder & CTO
Mar 2021–Sep 2022

Co-founded a B2B edtech startup focused on reskilling and upskilling the workforce. Delivered the flagship product sold to clients like AB InBev and helped close two funding rounds, the latter valued at $12M. Recruited and led the engineering and product founding team, and built the architecture to serve a web app, browser extension, and enterprise APIs.

FL4KSoftware Engineer
Sep 2020–Mar 2021

Worked on a platform helping primary school kids learn new languages. Built the front-end of the learning management system, designed and implemented digital educational games, and deployed the company's landing page.

Embaixadores da EducaçãoSoftware Engineer · Haas Center Fellow
Jun–Aug 2020

Joined an education NGO teaching entrepreneurship to low-income youth in Brazil. Conducted user discovery for the Empower Challenge product — interviewing users, creating prototypes, and iterating through usability tests. Led a volunteer team of 12 developers to deliver the platform in three months. The product launched at an event with 90,000+ students, attracting thousands of users in its first hours.

LinteProduct Designer (UI/UX)
Feb–Sep 2019

Designed the user interface of a legal tech SaaS product sold to over 70 companies including Accenture, Renault, and Sony. Conducted user research and usability tests, led ideation sessions, and delivered a design system along with documentation of the company's design process.

Volunteering

Brazil at Silicon Valley ConferenceVP of Communications
2020
SuperMentorCo-founder & Advisor
2019

Founded SuperMentor to democratize information about applying to U.S. colleges for low-income students in Brazil. Assembled the founding team and implemented the management and technology systems. The platform reached over 100,000 users.

Education

Stanford UniversityM.S. Computer Science, AI
2026
Stanford UniversityB.S. Symbolic Systems, AI
2025

Projects & Papers

Vision-based landing system diagram
Predictive Uncertainty for Runtime Assurance of a Real-Time CV-Based Landing System
Efficient neural architecture with calibrated uncertainty estimates for vision-based aircraft landing, using spatial Soft Argmax and adapted RAIM for fault detection.
DASC 2025 · Best of Session Award
arXiv →
Cluster-based rubric system diagram
Confiante G1: A Foundation Model for Language Learning Classes
AI foundation model that generates English language lessons at scale using reinforcement fine-tuning with a novel cluster-based rubric architecture, serving 1.8M learners across Latin America.
BeConfident Labs →
Knowledge neglect experiment results
Learning Errors from Context: Knowledge Neglect in Language Models
Tested whether LLMs exhibit knowledge neglect — a cognitive bias where reading fictional stories with embedded false facts leads to reproducing those errors on later knowledge tests. Gemma 3 1B's accuracy dropped from 96.6% to 18.9% after exposure, mirroring the human phenomenon including the illusion of prior knowledge.
Coming soon