One photo through two frozen towers. DINOv2 separates the bird as one region; Muse Glimmer shows no subject.
Vision Tower Bench
2026A benchmark for evaluating vision head capabilities
Immanuel PeterI work on multimodal and physical AI.
- Work
- MTS Intern at Tensormesh
- Study
- CS, Physics, & Math at UChicago
- Interests
- Vision, Multimodal, World Models, Physical AI
- Current
- Vision Tower Bench
- Hugging Face
- 20,096 downloads · 19,002 on datasets, 1,094 on models
Work
A frame from the CARLA autopilot dataset AutoMoE trains on, with the recorded speed, throttle, brake, and steering.
AutoMoE
2025A MoE self-driving model in PyTorch
Experience
Tensormesh
Mar 2026 – nowMember of Technical Staff Intern · Foster City, CA
- Contributed upstream to LMCache, adding a coordinator endpoint for listing pinned cache entries and per-key access tracking in the key directory.
- Shipped 35 PRs and 100+ commits across 9 company repos spanning LLM observability and inference optimization.
+ More− Less
- Integrated Arize Phoenix into Tensormesh’s observability stack, emitting OpenInference LLM traces from the router for all requests, and shipped SDK and CLI tooling for inspecting traces and spans.
- Reworked the Prometheus metrics for TTFT, throughput, and other rollups that power the serverless observability dashboards.
- Built an orchestrator microservice that pins prompts in LMCache asynchronously during inference, and exposed pin management through the API, SDK, and CLI.
- Built a KV cache discovery interface on top of the LMCache Coordinator API to browse cached keys, tokens, and utilization across engines.
- Led an internal research effort on key weighting for long-context inference, from evals to writeups.
Quantum Rings
Jun – Aug 2025Software Engineer Intern · Chicago, IL
+ More− Less
- Delivered 19 PRs, 43 contributions, and 15 completed GitHub issues across the internship, adding ~15K LOC and removing ~3.6K LOC while reviewing code and driving schema refactors.
- Migrated execution data from the user entity to a dedicated relational table with FKs, modularizing schema and ensuring test suite stability with no performance regression.
- Implemented a telemetry aggregation background worker (AWS SQS + TypeORM) to asynchronously roll up user execution activity, improving scalability and simplifying downstream analytics queries.
- Designed and deployed queue-driven execution processing to decouple heavy telemetry operations from the API, reducing request latency and enabling horizontal scaling.
- Built full-stack admin analytics dashboards with NestJS, Next.js, and Recharts, integrating SQL time-bucket aggregation and timezone-safe filtering to track user growth, active usage, and execution volume.
Open source
Education
University of Chicago
2024 – 2028B.S. Computer Science, B.A. Physics, B.A. Mathematics