Skip to content

Sidharth Kudupudi · Minneapolis, MN

Hi, I'm Sidharth.

I'm an engineer who likes turning messy real-world data into things that work, and then checking, honestly, how well they work. I recently graduated with an M.S. from the University of Minnesota, and I'm usually debugging something that worked perfectly yesterday.

Open to full-time engineering roles

Loading LiDAR sweep…

nuScenes · scene-0061 · LIDAR_TOP

— points · — annotated objects

vehiclepedestrian
A real LiDAR sweep from nuScenes, rendered live in your browser with its annotated 3D boxes. Drag to orbit. Data © Motional, CC BY-NC-SA 4.0.
infrared detect-and-track per frame

1.7 ms

infrared detect-and-track per frame

after my C++/CUDA/TensorRT port

AUROC for cobot fault detection

0.944

AUROC for cobot fault detection

at a 1.7% false-alarm rate

mAP50 across 5 folds

0.963

mAP50 across 5 folds

surgical tools, 0.87 ms TensorRT

projects with code and notebooks

10

projects with code and notebooks

every number traceable to a run

01Selected work

Things I've built, and how well they actually work

Problem, contribution, stack and a verified result for each. Every number comes from an executed notebook or a logged benchmark.

All 10 projects

01Computer Vision · Edge Deployment · Autonomous Systems

Infrared small-target detection & tracking

Problem
A 3–7 px target, 10–16 grey levels above cloud and city clutter, has to be found on every frame.
What I did
I built a detector and tracker for 3–7 pixel aerial targets in infrared video, then ported it from Python to C++/CUDA/TensorRT: 1.7 ms per frame, 9.3× faster, with tracks identical to the prototype.

1.7 ms

per 1280×1024 frame, enhance → detect → track

C++/CUDA/TensorRT FP16 on an RTX 5080

  • PyTorch
  • YOLOv8
  • CUDA
  • TensorRT
  • C++17
  • Eigen

02Robotics · Signal Processing

Physics-grounded fault detection & diagnosis for a cobot

Problem
A cobot repeats one pick-and-place cycle; collisions, lost cans and friction have to be caught without labelled faults to train on.
What I did
I built an interpretable pipeline on real cobot telemetry: an unsupervised detector flags faulty cycles at AUROC 0.944 with a 1.7% false-alarm rate, and a Random Forest names the fault.

0.944

AUROC, faulty vs normal cycles

1.7% false alarms on unseen normal cycles

  • pandas
  • scikit-learn
  • PCA
  • Random Forest
  • NumPy
Faulty cycles flagged at the training-derived limit
Faulty cycles flagged at the training-derived limit
Invalid position100%
Missed can100%
Motor commutation97.8%
Collision: carton86.4%
Wobbling station81.1%
Lost can66.2%
Collision: cable54.2%
Can weight52.5%
Axis weight42.3%
Collision: foam38.9%
Axis friction38.2%
Entangled20%
Normal cycles (false alarms)1.7%

Faults that change the cycle's dynamics strongly are easy; small extra friction or load is subtle.

Replay real recorded cycles in the interactive Lab

03Sensor Fusion · Autonomous Systems · Computer Vision

Camera–LiDAR fusion on raw autonomous-vehicle data

Problem
Segmentation says what is in the image; only LiDAR says how far away it is.
What I did
I built LiDAR-to-camera projection for a 120° f-θ lens from raw NVIDIA AV files and gave every segmented object a metric range, over two 605-frame drives.

2 × 605

frames fused end to end

city and highway clips

  • Python
  • NumPy
  • DracoPy
  • rosbags
  • ROS 2
  • YOLO11-seg
C++ output: Bent cable wire: ground-truth outline versus the C++ application's maskGround truthC++ output

04Computer Vision · Deep Learning · Edge Deployment

Industrial defect segmentation with a C++ TensorRT app

Problem
Defects must be localised to the pixel, on parts the model never saw, without rejecting good parts.
What I did
I trained YOLO11-seg on MVTec AD to 0.82 mask mAP50, wrote the C++ TensorRT app that serves it, and found that rebalancing VisA's data lifted mask mAP50 from 0.25 to 0.41 where a bigger model did nothing.

11 / 15

held-out defects located by the C++ app

mask IoU ≥ 0.5, one random case per category

  • Ultralytics YOLO11-seg
  • TensorRT
  • C++17
  • CUDA
  • OpenCV
  • ONNX

05Signal Processing · Computer Vision

Sub-pixel Fourier registration for wafer metrology

Problem
Overlay metrology needs image shifts measured to a hundredth of a pixel.
What I did
I implemented phase correlation with matrix-multiply DFT upsampling in NumPy: 1/100-pixel shifts that match scikit-image exactly on 50 of 50 tests, benchmarked across four wafer patterns and 15–40 dB of noise.

0.015 px

mean error on a DRAM array

25 dB SNR, 50 random shifts

  • NumPy
  • SciPy FFT
  • Phase correlation
  • scikit-image (cross-check)

Wafer pattern

Referenceclean design pattern
Targetshifted by (3.427, -5.184) px, 30 dB
Correlation peak, 1/100 px gridupsampled DFT
Residual after realigning|reference − target shifted back by the estimate|
Truth
(3.427, -5.184)
Estimate
…
Error
…
Compute, your browser
…

Live: drag the shift or noise and the algorithm re-runs in your browser.

02Interactive lab

Replay a cobot's cycle and watch the detector decide

Thirteen real cycles recorded from a collaborative robot. Pick a fault, play the cycle, and see the physics residuals leave the normal band before the detector and classifier give their verdict.

All demos

Loading 13 recorded cycles…

Real
Joint angles, residuals and detector/classifier outputs for 13 recorded voraus-AD test cycles (one normal, one per fault: the median-scoring cycle of each category).
Schematic
The arm drawing uses generic link lengths and axes 1–4 only. It shows the recorded motion, not the robot's real geometry.
Runs where
Pre-computed by the project notebook; your browser only replays and plots it. No model runs here, and nothing is uploaded.

03Technical depth

What the experiments taught me

Short engineering notes on findings that changed a design, and every latency figure on this site with how it was measured.

All notes

Latency, as measured

SystemMeasuredLatencyInputConditions
Infrared pipeline (C++/CUDA/TensorRT FP16)enhance + detect + track1.71 ms1280×1024mean per frame, decode excluded, RTX 5080
YOLO11n surgical-tool engine (TensorRT FP32)inference0.87 ms640×640mean of 3,451 runs, p99 1.03 ms, trtexec
MobileSAM image encoder (TensorRT FP32)encoder only4.31 ms1024×1024mean of 698 runs, p99 4.38 ms
YOLO11x-seg defect engine (TensorRT)inference6.7 ms1024×1024mean, p99 7.2 ms, trtexec
YOLO11s person tracker loopper frame, VLM idle → running17.9 → 20.2 ms p50dashcam videop95 20.4 → 296 ms on a shared GPU
Qwen2-VL-2B intent callper call824 ms p50clip + promptPyTorch, ≤ 40 new tokens
Sub-pixel registration (NumPy/SciPy)per image pair~24 ms median512×512CPU, κ = 100

Rows are not comparable with each other: models, inputs and measurement scopes differ. GPU rows: one RTX 5080.

04About

A bit about me

I'm an engineer with an M.S. in Robotics from the University of Minnesota, originally trained in mechatronics. I like the whole arc of a problem: making sense of messy data, building something that works, then making it fast and dependable enough that someone else can rely on it.

I care about evidence more than demos. I test on data a system has never seen, put the unflattering number next to the flattering one, and write down what didn't work.

More about me

Education

  • M.S. RoboticsGraduatedUniversity of Minnesota
  • B.Tech MechatronicsManipal Institute of Technology

Core stack

  • Multi-object tracking
  • LiDAR–camera fusion
  • C++17
  • CUDA
  • TensorRT
  • Kalman filtering
  • ROS 2
  • Isaac Sim
  • PyTorch
  • LoRA / QLoRA

Contact

Got an interesting problem? Let's talk.

Open to full-time engineering roles. If the problem is hard and the data is messy, I'm interested.