Swapnil Bhole

Projects

Grouped by area, most relevant first. Figures are measured results; details are on my resume.

Performance & ML Systems

Speed Demon 2026: four performance rewrites in 72 hours

Nov 2026 · Solo · C++20 · Code and optimization logs · Track D write-up

  • Top 3 Track D leaderboard
  • ~850× Track D vs. reference
  • 4/4 tracks entered
  • A 72-hour championship where every entrant gets the same task, the same data and the same scoring machine (Intel i5-13500, 12 threads, 8 GB cap). Output must match the reference byte for byte; the score rewards wall time, peak memory and short code.
  • Track D, graph traversal (top 3 on the leaderboard): 2,000 shortest-path queries on a road network with 1.74M nodes and 5M edges, from 72.69 s and 893 MB down to 0.085 s and 57 MB. I contracted pass-through chains, renumbered nodes with locality-preserving BFS so nearby roads sit together in memory, and used a bidirectional Dijkstra with a bucket queue. I wrote up how Track D came together, including the changes that made it slower, in 23 Minutes Before the Freeze.
  • Tracks A–C: AVX2 SIMD parsing, carry-less-multiply quote masks for CSV, AES-NI hashing, lock-free shared hash tables and branch-free percentile selection.
  • Every version checked against the reference output (sha256) and run under AddressSanitizer, UBSan and ThreadSanitizer.
TrackTaskPython referenceMy best runSpeedup
A · Ingest3.1M log lines → p50/p95/p99 latency and error rates per endpoint per hour32.05 s0.020 s, 24 MB~1,600×
B · Index50,000 boolean and phrase queries over a 250 MB corpus56.35 s, 2.2 GB0.217 s, 41 MB~260×
C · JoinCSV join with per-country totals and distinct counts14.44 s, 367 MB0.024 s, 9 MB~600×
D · Traverse2,000 shortest paths on a 1.74M-node road graph72.69 s, 893 MB0.085 s, 57 MB~850× · top 3

C++20 · AVX2 · lock-free data structures · PGO · perf · Valgrind · Docker

CUDA acceleration of CNN inference

2025 · University of Michigan

  • 355× speedup
  • 13.5 s → 0.038 s on a Tesla V100
  • Cut a two-layer CNN forward pass from 13.5 s to 0.038 s on a Tesla V100, 3.4× past the target.
  • Nsight-guided shared-memory tiling, thread coarsening, register tiling and kernel specialization.

CUDA · Nsight · C++

GPU acceleration of a configuration-novelty metric for ML interatomic potentials

2025 · University of Michigan

  • Custom CUDA kernels to accelerate a configuration-novelty scoring system used to improve machine-learned interatomic potentials, mapping the scoring onto a parallel reduction.

CUDA · C++

Assistive vision system for the visually impaired

2025 · Presented at the 4th Annual Disability Visibility in Engineering Symposium, U-M

  • 98.3% validation accuracy
  • End-to-end SSD + Inception-v3 object-detection pipeline with real-time audio feedback about nearby objects.
  • Designed the preprocessing, training and evaluation workflows.

PyTorch · SSD · Inception-v3 · text-to-speech

5G / Wireless PHY

5G O-RU PHY testbed on Zynq-7000

2021–2024 · C-DOT, Senior Engineer

  • 245.76 Mbps IQ stream
  • 1.6M+ packets, zero drops
  • 30.72 MB/s deterministic FPGA–ARM transfer
  • Built and integrated a 5G open radio unit (O-RU) PHY testbed on Zynq-7000 FPGA/ARM with Intel FlexRAN, supporting end-to-end 5G NR system validation.
  • Debugged spectral regrowth and PHY signal-quality degradation by correlating FPGA ILA captures with MATLAB reference models; isolated a fixed-point truncation in the signal-processing chain.
  • Root-caused packet loss under a 245.76 Mbps IQ stream to lwIP buffer exhaustion; reworked buffering to carry 1.6M+ packets with zero drops.
  • Implemented deterministic FPGA–ARM data movement with shared-memory ping-pong buffers and AXI DMA, sustaining 30.72 MB/s.

Zynq-7000 · ARM Cortex-A9 · Vivado ILA · AXI DMA · PetaLinux · MATLAB · Intel FlexRAN

LTE eNodeB 3GPP conformance and RF/PHY validation

2018–2021 · C-DOT, Engineer

  • Rel 9–12 3GPP
  • S1 · X2 · Uu interfaces
  • Developed and executed 3GPP LTE conformance and system-validation tests across S1/X2/Uu, covering FDD/TDD, 2×2 MIMO, VoLTE and NB-IoT.
  • Characterized and debugged EVM, ACLR, sensitivity, BLER, throughput vs. SNR/RSRP, TX power, spectrum and constellation with Keysight UXM, Amarisoft and TM500; root-caused RF/PHY performance and stability failures.

Keysight UXM · Amarisoft · TM500 · Python · C · Wireshark

ML-based joint coded modulation for LoRa

2024 · University of Michigan

  • SF7–SF9 baseline
  • 10% phase noise
  • Ported a deep-learning joint coded modulation framework from Bluetooth Low Energy to LoRa chirp spread spectrum.
  • Characterized the SF7–SF9 baseline at 1e-4 BER and 16 dB SNR, and matched it with a GRU codec under 10% phase noise.

Python · PyTorch · DSP

RTL & Computer Architecture

2-way superscalar out-of-order RISC-V processor

2025 · University of Michigan

  • 2-way superscalar
  • R10K register renaming
  • Designed and verified a complete out-of-order CPU in SystemVerilog with R10K-style register renaming, branch prediction and L1 instruction/data caches.
  • Ran the verification test suites with Synopsys VCS and debugged in Verdi.

SystemVerilog · Synopsys VCS · Verdi

Duo-binary turbo decoder for DVB-RCS

C-DOT · DRDO-funded

  • RTL turbo decoder block (fixed-point Max-Log-MAP, duo-binary) for satellite return channels, delivered as a revenue-generating IP for C-DOT.

Verilog · RTL · turbo codes

Embedded & Products

OpenExo data-collection software

Sep 2026 – present · Software Lead, University of Michigan

  • Lead the data-collection software for a wearable exoskeleton research platform.
  • Extend the open-source OpenExo C++ firmware (Teensy) and a Raspberry Pi Pico W sensor platform to stream IMU, EMG and FSR data to a Python/Qt GUI.

C++ · Teensy · Raspberry Pi Pico W · Python · Qt

SIM908 fuel-notification device

2019 · Commercialized by TSecond, India

  • 5 m GPS accuracy
  • Embedded device that uses pulse-counting flow sensing and GPS/GSM telemetry to detect and report fuel malpractice. Commercialized by TSecond, India, which patented the product.

AVR · SIM908 · GPS/GSM · PCB design

ESP32 IoT telemetry and control

2022

  • Real-time ESP32 telemetry node with a cloud relay and a React Native control app: firmware, transport and app.

Embedded C · ESP-IDF · MQTT · React Native

SmartQRX: tamper-detecting QR code

2025 · Semifinalist, Michigan Business Challenge

  • A QR code with visible color-change indicators for moisture, temperature and tamper events in supply-chain packaging; the code stays scannable.

Product design · materials · supply chain

Other projects

Quantum denoising of single-cell RNA sequencing data

2025 · University of Michigan

  • Applied quantum state tomography to scRNA-seq denoising; outperformed classical MAGIC imputation on the PBMC dataset.

Qiskit · Python

PathForward: graduate career platform

2026 · Semifinalist, Munger Case Competition

  • A platform connecting current graduate students with alumni on similar career paths. Built the MVP and the business plan. Repository.

React · Node