Hi, I'm Arthur.

Large-scale applied ML across post-training, agent systems, in-context learning, and graph-based representation learning. Core areas: retrieval / search / clustering, representation learning, GNNs, agents, post-training, and RL.

• Agent systems, RL & imitation learning. Built end-to-end LLM agents for production workflows (sequential, parallel, sub-agent, critic-agent designs) and large-scale RL / imitation learning experiments on 10B+ user-interaction trajectories, spanning offline and online settings.
• Post-training & ICL. Fine-tuned Gemini encoder models via SFT pipelines in JAX for downstream tasks; improved agent quality with retrieval over multimodal embeddings and vector stores for few-shot ICL.
• Infrastructure & scale. Operate ML systems at Google-scale: petabyte-scale graph and feature pipelines with second-level latency; distributed training in JAX on TPU pods; online ANN serving over 1B+ learned vectors; production data flows on Pub/Sub, Kafka, and Spanner. Comfortable owning the path from raw event streams through training, evaluation, and online serving.
• Representation learning & retrieval. Built GNN-based representation learning on heterogeneous graphs with 50B+ edges, with ANN retrieval over 1B+ learned vectors for downstream agent and recommendation surfaces.
• Agentic coding influence. Have been using and promoting the Plan–Define–Act workflow for agentic coding since May 2025; authored an internal doc on Gemini CLI and agentic coding that reached 12k+ unique views within Google/Alphabet, and delivered tech talks on GenAI developer tooling.

ML systems for large-scale financial time-series, focused on anomaly detection, feedback-driven model improvement, and end-to-end production ML pipelines.

• Designed large-scale anomaly detection workflows for financial time-series in collaboration with Portfolio Managers and Quants.
• Built a feedback-loop system to address distribution shift and pattern discovery, iteratively improving model accuracy and stability.
• Engineered end-to-end ML pipelines: data processing, feature engineering, training, tuning, evaluation, versioning, and serving.

HCI and human-in-the-loop ML research on multimodal interaction, wearable input, and human-drone control — focused on subtle one-handed drone interaction using touch, force, and IMU sensing.

• Designed and evaluated four multimodal control schemes for one-handed drone interaction, identifying a method that outperformed a two-handed commercial baseline on task completion time.
• Built a ring-sized wearable controller enabling joystick-like drone navigation through subtle thumb interaction and wrist pitch/roll gestures.
• Conducted user studies on task completion time, perceived workload, and interaction footprint to derive design guidelines for compact multimodal interfaces.

CV research and engineering for visual recognition on edge devices — dataset construction, transfer learning, model compression, and deployment to mobile and embedded platforms. Worked on car model classification with Inception-v3 transfer learning.