Nathan Lemma
Experience2023 – present

Roles and teams

Research at UT Austin, three teams at Caterpillar, and extras.

  • UT Austin Research Undergraduate Research Assistant

    Mar 2024 - Ongoing

    • Machine Learning for Systems · LDOS Lab · Akella

      Aug 2026 - Ongoing

      • Building a harness that streams live vLLM scheduler and memory-manager stats to an LLM agent, which tunes vLLM's load balancer across replicas.

    • Systems for Machine Learning · SysML Lab · Yadwadkar

      Aug 2025 - May 2026

      • Rebuilt 4 state-of-the-art GPU-sharing inference serving systems (GPULets, Usher, FGD, ParvaGPU) to benchmark EnerTune.
      • Debugged and ported each baseline's research prototype onto our 16-A100 cluster, enabling the 1.4-2.3x lower energy result.
      • Profiled the baselines to show EnerTune's analytical model cuts profiling time 7.3x on average, up to 17.3x, versus prior systems' profiling.

    • Computer Architecture & FPGAs · FAST Lab · Chiou

      Mar 2024 - Dec 2024

      • Built an instruction-accurate RISC-V simulator in C++ so the team could test the Primate compiler without slow FPGA synthesis.
      • Designed I/O instruction handling and compiler-checking tools that sped up developing the hardware and software side by side.

  • Caterpillar Inc. Software Engineering

    Summer 2024 - Ongoing

    • Data Acquisition & Analytics Team

      Summer 2026, part-time ongoing

      • Building a tool attributing cellular data costs across 20,000+ Caterpillar machines, raising cost accuracy 34% on average and up to 55%.
      • Processing a 100 GB daily load from a 30-50 PB pipeline to replace whole-SIM-bill accounting with per-megabyte attribution.
      • Deploying a nightly AWS Lambda that calls carrier APIs to locate each machine, aggregate usage, and compute its actual cost.

    • System Diagnostics Team

      Summer 2025, part-time through 2025-26

      • Wrote hardware-in-the-loop tests covering 30 I/O types (PWM, analog, digital) to set a baseline for the whole diagnostic suite.
      • Automated key steps in the firmware build process used daily by 15+ engineers, cutting their configuration time by 10%.

    • Platform Software Team

      Summer 2024, part-time through 2024-25

      • Designed hardware-abstraction (HAL) and API layers for engine control modules (NVM, analog, digital I/O), now in production.
      • Developed firmware in C and Python that cut new I/O support time by 15%, verified via .elf memory-layout checks.

  • Longhorn Racing (Formula SAE) Battery Management Developer

    Aug 2023 - May 2025

    • Led design of the car's battery monitoring system and wrote RTOS firmware modules handling drivers, I/O and scheduling.
    • Built the State-of-Charge algorithm that cut the driving-range margin of error to under 5%.