Amil Agrawal

I study computer science and math at Brown University. I’m interested in reinforcement learning, world models, and efficient inference.

At Brown

Internships

Snap Inc.

ML Research Engineer Intern

Applied research on L1 recommendation foundation models for 500M+ Snapchat user feeds.

AMD

ML Engineer Intern

Inference infrastructure optimization and kernel-agent post-training for ROCm.

RTX

Robotics Engineer Intern

Multi-agent reinforcement learning for collaborative aerial autonomy.

Bloomfilter

Software Engineer Intern

First SWE intern, working on AI and core product systems.

RSGbot

A Minecraft speedrunning agent.
Screenshots in, keyboard and mouse actions out.

RapidTinyRec

A CUDA and Triton engine for MiniOneRec, a 1.5B generative recommender. It replaces beam search with two passes that bound scores and certify the top K recommendations. Learned bridges let item-code tokens skip early layers.

Median request latency

Office · H100 · FP16 · batch 1
Same engine and scoring rule

  • Beam-505.13 ms
  • Certified, full depth4.21 ms
  • + third-code bridge3.06 ms
  • + first-level bridge2.61 ms
2.7× throughput
970 → 2,624 users/s at batch 64 versus beam-50 in the same engine.
4.5× faster than FlashRec
2.57 ms versus 11.48 ms for the BF16 FlashRec Qwen2 port in a separate run on the same H100, prompts, and catalog.
18.25% HitRate@10
Adding SASRec scores during search raised hit rate by 1.36 percentage points at 2.83 ms per request.

The late-entry decoder certified a top-10 list for all 4,866 Office test users under its own scoring objective. HitRate@10 was 16.87% versus 16.81% for beam-50. The top item agreed with full-depth decoding for 79.5% of users.

Earlier A6000 benchmarks

Original → optimized FP32

868.8→114.8ms median

7.6× median speedup. Same rankings; separate recorded runs.

Original loop · FP32Optimized FP32
Loading recorded requests…

Original → experimental BF16

868.8→56.5ms median

15.4× median speedup. Rankings change; separate recorded runs.

Original loop · FP32Experimental BF16
Loading recorded requests…
Other comparisons & measurement notes

Overall comparisons align the original released-loop timings with later optimized timings by cohort index. Same Office checkpoint, 500 users, RTX A6000, batch 1 and beam 50; measured in separate runs, not a new paired experiment. Speedups are ratios of the two measured medians, not multiplied per-step gains. The optimized FP32 run matches all 500 original ranking lists; scores are not bitwise identical. BF16 is experimental and changes rankings. Model loading, compilation, capture and warmup are excluded. Traces show 80-request windows in cohort order and retain all measurements. The individual optimization comparisons keep their own paired controls; the isolated BF16 comparison marks its 693.8 ms outlier above the 90 ms axis. Original timings · FP32 evidence · Precision & quality

SysVRC

Autonomous VRC routines in ROS 2 and Gazebo. The field, controller estimate, and command trace on one timeline.

On the field

6121C · Spin Up

Autonomous skills
Driver skills

PAVER

Panoramic Alignment and View-based Evidence Recovery

Recovering occluded street-view pixels from neighboring panoramas, using feature matching and geometry.

Before: the capture car obscures road markings in Strasbourg After: partial road recovery using observed pixels from neighboring panoramas
Strasbourg · capture-car removal
Before: a controlled foreground mask beside a Beijing boulevard, with buildings and the road visible around it After: saved donor-view recovery of the masked foreground in the same synthetic Beijing boulevard scene
Beijing · controlled occlusion
Quai Saint-Nicolas

A · find the visible road pixels

Reconstructing Mercury

Reconstructing Mercury’s southern gravity field with lunar transfer learning and topography.

Input · degree 25

Gravity · mGal
Gravity color scale: blue for negative, white for zero, red for positive
−2000+200
Mercury’s gravity field at spherical harmonic degree 25

wild-env

A wildfire-response environment for multi-agent learning: fire spread, aircraft dispatch, suppression, and logistics.

stemp

Four stems, twelve pads, a step sequencer.
A music instrument for macOS.