0
Downtime migrating Uber's Ray training controller across 2,000+ ML pipelines to regional federated compute.
MLE · ML Systems
Felipe Felix Arias
MLE on Uber's ML Training team. I build GPU training infrastructure, GenAI systems, and support LLM fine-tuning — spanning infra, applied ML, and production deployment.
Founder of Marovi AI, a research comprehension and translation platform that makes knowledge accessible across languages. Former NSF Graduate Research Fellow at UIUC.
Marovi AI
Research comprehension & translation platform
NavIRL
Agentic indoor social-navigation toolkit
ML Training Infrastructure
Evaluation & Data Pipelines
Agentic Systems
Distributed Training
Ray / Kubernetes
Python
PyTorch
Go
San Francisco · Fluent in English and Spanish · Born in La Paz, Bolivia
Impact at a Glance
Signals from production systems, research, and Marovi AI.
3.7%
Drop in unsafe Uber driver behavior from a GenAI + telemetry-based distracted driving detection system.
Top 10% x2
Y Combinator selections for Marovi AI, a research comprehension and translation platform across 9+ languages.
4 Labs
Research at Stanford, UC Berkeley, UIUC (NSF fellow), and published with Google Brain (ICRA 2021).
Technical Focus
Systems areas I work in.
ML Training Infrastructure
Job controllers, GPU scheduling, training platform reliability, and supporting LLM fine-tuning (LoRA, QLoRA, full) at scale.
Evaluation & Agentic Systems
Agentic systems, RAG, structured evaluation, and production GenAI with safety guardrails.
Distributed Training Infrastructure
GPU cluster orchestration, Ray on Kubernetes, DDP/FSDP, and federated compute at scale.
Research Comprehension & Translation
Crowdsourced translation and AI-powered learning platforms that make research accessible across languages.
Experience Highlights
Selected work across LLM systems, GenAI, and research.
Uber — ML Training Team
- Build job controllers and GPU training infrastructure; led the move from zonal to regional federated compute with zero downtime.
- Co-designed Ray job submission across federated Kubernetes clusters and shipped AI safety systems plus internal assistants for incident response.
- Leading mTLS rollout for batch workloads and serving as primary point of contact for LLM fine-tuning partnerships.
Marovi AI
- Building a research comprehension and translation platform with a provider-agnostic API across OpenAI, Anthropic, Google, and DeepL.
- Designed crowdsourced correction loops and agentic translation pipelines; content across 9+ languages; selected top 10% at Y Combinator twice.
- Full-stack: AWS deployment, MediaWiki frontend, Python orchestration, Pydantic schemas, and benchmark evaluation (FLORES, XL-SUM).
Research — Stanford, UC Berkeley, UIUC, Google Brain
- UIUC Parasol Lab (NSF fellow): Self-supervised data labeling via simulation for multi-agent motion planning (5x speedup). Published with Google Brain at IEEE ICRA 2021. Master's thesis on motion pattern prediction.
- Stanford Hazy Research: Extended Snorkel's programmatic data labeling for multi-sentence relation extraction (+12% F1). Collaborated with Alex Ratner (now CEO of Snorkel AI) and Christopher Ré.
- UC Berkeley: Built shake-gesture detection for E-mission, an open-source mobility platform (+7% F1 via SVM on imbalanced signals).
Selected Projects
A few representative projects.
Marovi AI: Research Comprehension & Translation
Research comprehension and translation platform with crowdsourced corrections, a provider-agnostic LLM API, and agentic translation pipelines across 9+ languages.
NavIRL: Indoor Social Navigation Toolkit
Agent-driven indoor social-navigation toolkit with deterministic simulation, verification gates, and an Aegis overseer pipeline for VLM-backed qualitative realism checks and tuning.
Avoidance Critical Probabilistic Roadmaps
Self-supervised programmatic data labeling through simulation: learns avoidance-critical regions from generated collision data and distills them into spatial models for scalable multi-agent motion planning.
Programmatic Data Labeling for Multi-Sentence Relation Extraction
Used Snorkel's programmatic labeling framework to generate training data for cross-sentence relation extraction via LSTMs. Achieved 12% F1 improvement through novel multi-sentence labeling functions and multi-task learning.
Video as a Sensor for Urban Risk
Weak supervision and behavioral heuristics surface risky roadway events when labeled data is scarce.
Publications
Selected academic work.
Now
Building ML training infrastructure and GenAI systems at Uber. Growing Marovi AI on the side. Experimenting with agentic research via NavIRL.
Outside Work
I play guitar, drums, Latin dance, and travel. Covers and clips on Instagram.