CS & Physics @ University of Chicago · ipeter.dev

I work on multimodal and physical AI: large-scale training, linear algebra, and inference systems.

• MTS Intern @ Tensormesh, working on GPU management, LLM observability, and KV cache optimization for inference infrastructure.
• Deepening into distributed training and multimodal architectures

• LMCache: added coordinator APIs for listing pinned cache entries and tracking per-key access counts.
• vLLM Production Stack: added router volume mounts for read-only root filesystems and fixed the default NVIDIA runtime class.
• Pyrefly: standardized the Rust error-summary module and imports from summarise to summarize.
• Brev CLI: authored the core rsync-first file transfers with automatic SCP fallback, preserving incremental transfers for repeated copies.

• Vision Tower Bench compares frozen vision towers across recognition, geometry, cross-view correspondence, and occlusion to guide tower, layer, and projector choices before multimodal training. Code · Towers
• AutoMoE is a 53M-parameter driving model with four perception experts, learned gating, and a trajectory policy. Trained experts on BDD100K and nuScenes, then adapted them to CARLA. Released 82K-frame multimodal and 68K-frame image datasets.
• Hostess compiles a declarative stack file into Kubernetes workloads, builds container images, provisions and wires Postgres and Redis, and creates isolated PR preview environments through a GitHub App.
• Redis Operator manages Redis pods and persistent volumes directly, with standalone, Sentinel, and cluster modes and fencing-first failover to prevent split-brain.
• Postplan provides an easy toolchain for agents to publish HTML artifacts and static assets through an authenticated curl requests.

ipeter.dev · Resume · LinkedIn · Email