Computer Vision · Perception · Robotics

NishanthVanipenta

Computer Vision & Perception Engineer

Camera + laser 3D measurement, from optics and calibration to GPU model serving. Systems I built scan 85,000+ tires a day across 1,200 drive-over systems and 8,000+ handheld scanners.

Resume ↗ LinkedIn ↗ GitHub ↗ Email

Dallas–Fort Worth, TX · open to relocate 2 Patents Pending TrailBlazer Award 2025

85K+

Tires scanned / day

1,200

Drive-over systems

±0.5mm

Tread accuracy

Patents pending

2025

TrailBlazer Award

01 / About

About me

Profile

I'm a computer vision and perception engineer who builds camera and laser measurement systems end to end: I choose the optics and sensors, calibrate them, write the geometry and deep-learning pipelines, and serve the models in the cloud. Most computer vision work sits on either the model side or the hardware side. I work on both.

I started in mechanical engineering at SRM University, got hooked on how machines could see, and did my M.S. in Robotics at the University of Maryland. On a General Motors program at ZD Metal Products, a Tier 1 supplier, I built vision inspection and programmed industrial robots for the Hummer EV frame line.

Since January 2022 I've led the computer vision behind TraXtion's tire-scanning products. Systems I built scan 85,000+ tires a day. Two of those inventions are patent pending, and the work earned TraXtion's highest honor, the TrailBlazer Award 2025.

My working rule: 1.01³⁶⁵ = 37.78. Get 1% better every day.

Beyond the Code

BaseDallas–Fort Worth, TX
RunningActive runner — targeting 10K & Half Marathon in 2026
TrekkingEverest Base Camp — goal 2027
SportsCricket, Volleyball, Swimming
WatchingFormula 1, Cricket

02 / Experience

Work experience

5+ years across perception, machine vision and industrial robotics

Jan 2022 — Present

Computer Vision Engineer

TraXtion

Dallas, TX

I lead the computer vision behind TraXtion's tire-scanning products: choosing cameras, lenses and lasers, calibrating them, building the measurement and deep-learning pipelines, and running them in production cloud services.

Built the production vision pipeline for a drive-over tread scanner on 1,200 systems scanning 70,000+ tires a day: multi-camera laser triangulation measuring tread depth on moving vehicles to ±0.5 mm of a manual gauge in about 15 seconds. Patent pending.
Selected the imaging hardware (industrial cameras, lenses, line lasers, optical filters, IMU and motion sensors) and ran field-of-view and resolution analysis, including a dual-camera rig for commercial trucks.
Made laser-based 3D measurement reliable outdoors, in rain, glare and low light, by combining optical filtering with robust point clustering.
Led field evaluation of a 6-camera structured-light 3D scanner; showed it could not meet accuracy outdoors and recommended the laser-triangulation approach that went to production.
Built an in-house system that reads tire sidewall text from any viewpoint, including phone photos, at 7.5× lower cost per image than the outsourced vendor with 66% more accurate detections. It became a new product, TireSpec. Patent pending.
Took Segment Anything (SAM) from a research script to a production GPU endpoint behind a feature flag with automatic fallback, raising detection rate from 40% to 95%+.
Built the sensor-fusion engine for a handheld laser tread scanner on 8,000+ devices scanning 10,000+ tires a day (0.01 mm resolution).
Built a ground-truth regression suite that gates every release; hardened CI/CD and production cloud services.
Guide technical work for a team of 5. TrailBlazer Award 2025, the company's highest award.

PythonC++OpenCVPyTorchCamera CalibrationLaser TriangulationSegment AnythingSensor FusionAzure MLDockerCI/CD

Products I built the vision for

▶ Demo

Drive-over scanning

TreadSpecX

Drive-over tread analytics that scans every tire as a vehicle rolls in, with DMS reporting, wear analytics and automated customer messages.

Laser TriangulationMulti-CameraCloud

Computer Vision

WheelSpec

High-speed optical capture of tires at 1–5 mph: rim, sidewall and license plate with no human interaction.

High-Speed ImagingOCR / ALPRReal-Time CV

Mobile Diagnostics

GrooveGlove 2.0

Portable wireless tire diagnostics with 4,000-point tread measurement and alignment diagnosis.

Laser ProfilingWirelessTread Analysis

AI Vision

DamageSpecX

Full-vehicle damage detection that produces a condition report in seconds from a multi-camera array.

Deep LearningObject DetectionMulti-Camera

Laser Scanning

GrooveSpec / GrooveSpecX

Wi-Fi (GrooveSpec) and Bluetooth (GrooveSpecX) handheld laser tread scanners for iOS and Android: 4,000-point profiles, license plate recognition and instant reports.

Laser ScanningSensor FusionWi-Fi / BluetoothiOS & Android

AI Web Plugin

TireSpec

A website plugin that reads a customer's tire sidewall from a phone photo and returns replacement matches with live pricing and booking. Built on my patent-pending sidewall-reading technology.

Computer VisionOCRMobile WebPatent Pending

Mar 2021 — Jan 2022

Robotics Engineer

General Motors program · via ZD Metal Products (Tier 1 supplier)

Petoskey, MI · Hummer EV

Before perception, robotics. I built vision inspection and programmed industrial robots for Hummer EV frame production, and learned how far a lab prototype is from a system that runs on a factory floor every shift.

Led a team of 3 to build a low-cost vision system, raised as a GM 8D corrective action after worn or wrong drill bits produced out-of-tolerance holes. It measures tapered-hole diameters on the Hummer EV frame against a 4.75–6.25 mm tolerance to 0.08 mm precision, raising defective-part detection by 14%. Built on Zhang camera calibration, undistortion, gradient edges and FFT-convolution circle detection.
Built an error-proofing station that confirms each assembled pillar matches its left- or right-hand receiver: a FLIR GigE camera with a 25 mm telephoto lens, AWS Textract OCR for serial numbers, circle detection for ID features and HSV thresholding to count screws and washers, multithreaded on the Spinnaker SDK.
Built presence checks for slider bearings, pins, buggy jumpers and bushings on Corvette and Hummer handles. Mask R-CNN took 14 s per image on CPU, too slow for the line, so I retrained a multi-class YOLOv5 detector in PyTorch (89% accuracy) and deployed it on an embedded device inspecting parts on the conveyor at 10–12 FPS.
Programmed FANUC and UR5e robots; diagnosed PLCs, pneumatics and motor controllers.

FANUCUR5eYOLOv5Mask R-CNNPyTorchCamera CalibrationFLIR / SpinnakerAWS TextractIndustrial Automation

Pillar assembly checked for ID and missing screws

Mask R-CNN segmenting pins and bushings on handles

YOLOv5 presence checks on the production line

03 / Case studies

Case studies

Problem, approach and result for four systems running in production

Laser triangulation · multi-camera

Measuring tread depth on moving cars

1,200drive-over systems

70K+tires / day

±0.5 mmvs manual gauge

~15 sper vehicle

Problem

Manual tire inspection is slow and inconsistent, so worn tires get missed. The goal: measure tread depth on every tire of a vehicle as it drives over a scanner, outdoors, without anyone touching the car.

Result

Runs unattended on 1,200 drive-over systems scanning 70,000+ tires a day, within ±0.5 mm of a manual gauge in about 15 seconds per vehicle. Patent pending.

Approach

Four cameras and line lasers with red bandpass filters fire on a hardware trigger as the vehicle rolls over. Each tire runs its own pipeline in parallel: rectify with calibrated homographies, stitch two camera views, extract the laser line with adaptive thresholding and a five-stage DBSCAN filter (density, size, spread, shape, covariance) with wet and dry modes, then measure grooves with a semi-global convex hull and peak detection, converted to millimetres. The four tires merge into condition codes, alignment and rotation advice computed locally, including dual rear wheels on trucks. An earlier six-camera structured-light design failed outdoor accuracy tests, so I recommended this approach instead.

My role

Led the vision pipeline end to end: hardware selection, calibration, measurement, integration and release packaging.

Trigger→4 cameras + lasers→Homography + stitch→Laser line (DBSCAN)→Convex-hull depth→Condition codes→Report

PythonOpenCVNumPyscikit-learnmultiprocessingBasler cameras

Sensor fusion · signal processing

Laser tread depth from a handheld scanner

8,000+devices

10K+tires / day

0.01 mmresolution

1–25 mmcar to truck tread

Problem

A technician sweeps a glove-mounted laser across a tire. Hand motion is unpredictable, raw data is noisy, and tread ranges from passenger tires to 25 mm truck tires. Groove depths and wear calls must arrive in real time, with no false readings reaching the technician.

Result

The production engine behind a scanner on 8,000+ devices scanning 10,000+ tires a day. Fewer false errors reach the field because laser and motion faults are flagged separately, including a crash I traced that had been misreported as a sensor error.

Approach

I fuse the laser profile with an optical-navigation motion sensor. Motion checks (speed, skew, drift, width, surface contact and direction) reject bad scans immediately with a specific error code that tells the operator how to rescan. Clean scans are calibrated to physical units and skew-corrected; the tread baseline comes from a convex hull, the profile is smoothed with Savitzky–Golay filtering, grooves are found by peak detection, and non-linear least-squares fitting fills sensor gaps. Depths pass an anomaly filter and feed a wear analyzer (center, shoulder, camber, toe); a four-tire service turns them into rotation and alignment recommendations. Truck tires are detected automatically.

My role

Designed, built and run the engine: sensor calibration, geometry and filtering, APIs, error taxonomy and monitoring.

Laser + motion sensor→Scan validation→Calibrate + de-skew→Convex-hull baseline→Groove detection→Depth + wear→Vehicle advice

PythonNumPySciPypandasAzure FunctionsSite24x7

OCR · computational geometry

Reading tire sidewalls from phone photos

7.5×lower cost / image

+66%more accurate reads

12,000models catalogued

150+image regression set

Problem

An outsourced OCR API was expensive and only worked with a fixed camera parallel to the tire. Phone photos, with any angle, glare and worn lettering, failed.

Result

7.5× lower cost per image than the vendor with 66% more accurate detections, from any viewpoint. It became a new product, TireSpec. Every extraction change is validated against a 150+ image ground-truth set before release. Patent pending.

Approach

Classical vision finds the tire and rim (Hough circles and ellipse fitting, with an edge-density score to pick the right candidate), then the curved sidewall is unwrapped from polar to flat coordinates so embossed text reads straight. OCR runs as a confidence-weighted cascade that re-runs perspective-corrected and raw passes only when brand or size is still unresolved. Every read is validated against a catalogue of 600+ brands and ~12,000 models with fuzzy matching and alias resolution, and DOT codes are decoded into manufacture dates. A Segment Anything detector now runs alongside behind a feature flag.

My role

Designed and built the pipeline, from research to production, and own its reliability.

Photo→Rim + tire detection→Polar unwrap→OCR cascade→Catalogue match→Brand · size · DOT

PythonOpenCVGoogle Document AIAzure FunctionsAzure SQLBlob Storage

MLOps · segmentation

Segment Anything, from research script to GPU endpoint

40% → 95%+detection rate

30 s → 0warm-up on live requests

0caller code changes

T4GPU endpoint

Problem

The hand-tuned classical tire detector failed at unusual capture angles. A text-prompted Segment Anything (SAM3) prototype did far better, but only as a research script on a laptop.

Result

Detection rate rose from 40% to 95%+, live traffic never pays the warm-up cost, and the calling service needed no code changes.

Approach

I ported it into a stateless scoring module on an Azure ML managed online endpoint: Docker image, infrastructure as code with GPU sizing, timeouts, concurrency limits and health probes. Its response matches the classical detector's ellipse geometry, so the calling service switches with a feature flag and falls back automatically on timeout or failure. The one-time CUDA warm-up moved into the startup hook, and a staged local test path (a Flask wrapper, then Azure's own inference server) let the other team integrate before the cloud endpoint existed.

My role

Owned it from prototype to production, including the API contract and rollout plan.

Image→Feature flag→SAM3 on GPU→Ellipse fit + gate→Pass / retake

PythonSAM3PyTorchAzure MLDockerIaC (YAML)

04 / Projects

Engineering projects

Agentic AI, research, autonomy and perception outside the day job

Agentic AI

Multi-agent business workflow

A multi-agent system for a paper-supply business covering inventory, sales, accounting transactions, customer support and lead generation. Built as my Udacity Agentic AI Nanodegree capstone. I also use Claude Code with skills, hooks and MCP daily, and run local models with Ollama.

Multi-agentTool callingMCPOllama

Research · UMD CEEE

RoCo: robot air-conditioner

As a research assistant at UMD's Center for Environmental Energy Engineering, I built RoCo's face recognition and tracking system on my own, covering design, hardware, software and testing. PoseNet raised tracking accuracy from 80% to 95%, with 91% recognition accuracy. It runs on a Raspberry Pi 4B with a Coral Edge TPU at 32–36 FPS and costs 40% less than the previous prototype.

TensorFlow LitePoseNetCoral Edge TPURaspberry PiOpenCV

Robotics · Autonomy

Maverick

A multi-agent autonomous vehicle system that navigates under constrained compute, with every decision made on the robot in real time.

ROSEdge AIPythonC++

Industrial Robotics

ARIAC 2020

Agile Robotics for Industrial Automation Competition: I built perception and manipulation pipelines for a dual-arm robot building kits in a simulated warehouse.

ROSC++MoveItGazebo

View code ↗

Computer Vision · 3D Geometry

Visual odometry

Recovers a car's camera trajectory from driving video. SIFT features are matched with FLANN, and a fundamental matrix is estimated with the 8-point algorithm inside my own RANSAC loop. The essential matrix comes from the camera calibration, and the correct pose is chosen by triangulation.

PythonOpenCVEpipolar GeometryRANSAC

View code ↗

Computer Vision · ADAS

Lane detection

Finds lane lines in day and night driving video. Each frame is undistorted, warped to a bird's-eye view with a homography, and filtered with blur, Canny edges and intensity masks. Probabilistic Hough lines are then fitted into left and right lanes.

PythonOpenCVHomographyHough Transform

View code ↗

Reinforcement Learning

Q-learning maze solver

An agent that teaches itself to solve 4×4, 5×5 and 10×10 mazes with walls and pits. It uses epsilon-greedy exploration and Bellman updates to a Q-table until it converges on the shortest safe path.

PythonQ-learningReinforcement Learning

View code ↗

05 / Skills

Core skills

Hardware × Software × Intelligence

Computer Vision

Camera calibration
Laser triangulation
Homography & 3D geometry
Structured light & point clouds
Segment Anything / segmentation
OCR
OpenCV

ML / Deep Learning

PyTorch
TensorFlow
YOLO
Mask R-CNN
scikit-learn
Model optimization

Deployment & Cloud

GPU model serving
Docker
CI/CD pipelines
Serverless APIs
Azure (ML, Functions, SQL)
AWS · GCP

Sensors & Hardware

Industrial cameras & optics
Line lasers & optical filters
IMU & motion sensors
Sensor fusion
Hardware triggering
Edge AI (Raspberry Pi, Coral TPU)

Robotics

ROS
FANUC
UR5e
MoveIt / Gazebo
PLCs & pneumatics

Languages & AI

Python
C++
SQL
Bash
Agentic AI (MCP, tool calling)
Claude Code · Ollama · RAG

06 / Education

Education

Mechanical roots × robotics wings

Certification

Udacity

Nanodegree — Agentic AI

2026 · Verified Certificate

Project-based training in building autonomous, multi-step AI agents: tool use, planning, memory and orchestration. Verify certificate ↗

Master's Degree

University of Maryland, College Park

M.S. Robotics

2019 – 2021 · College Park, MD

Vision-based autonomous systems, robot control and industrial automation, including the ARIAC competition in ROS and C++. Robotics research assistant at CEEE on RoCo, a robot air-conditioner, and graduate teaching assistant for Human-Robot Interaction.

Bachelor's Degree

SRM University

B.Tech. Mechanical Engineering

2014 – 2018 · India

Mechanical systems, manufacturing and control theory: the physical intuition behind building sensors and scanners that work in the real world.

07 / Contact

Let's talk

Building perception for robots, vehicles or sensors?
I'm always glad to talk about hard vision problems.

Email Me LinkedIn ↗ GitHub ↗ Blog Resume ↗