Ishaan Singh

sing1775@purdue.edu LinkedIn ↗ GitHub ↗ Résumé (PDF) ↓

I build computer vision pipelines and deploy models on edge hardware.

Seeking Summer 2027 internships in machine learning, computer vision, and embedded systems.

Core skills: Python · C · PyTorch · Linux · Docker · TensorRT

01 / EDUCATION

Education

Purdue University

Elmore Family School of Electrical and Computer Engineering

B.S. Computer Engineering
Concentration in Artificial Intelligence and Machine Learning

GPA: 3.54/4.00

Dean’s List · Fall 2024–Spring 2026

Relevant coursework

Completed: Introduction to Digital System Design · Advanced C Programming · Python for Data Science · Linear Algebra and Differential Equations

In progress: Data Structures · Signals and Systems

Expected May 2028
West Lafayette, Indiana

02 / PROJECTS

Engineering projects

Selected work in computer vision, machine learning, and embedded systems.

CrowdHuman-trained detector output on a street scene, showing predicted person boxes and confidence scores. — **Detection example**
CrowdHuman-trained model on a street-video frame. Boxes show predicted pedestrians and confidence scores; this is a qualitative example.
View full size ↗

COMPUTER VISION / EDGE AI

Pedestrian Detection & Edge Deployment

Code & evaluation results ↗

3.6× faster inference · 34.4 → 9.6 msJetson Orin Nano · 512 × 512 input · mAP50–95 maintained at 0.211

I trained six YOLOv8m detectors and compared how they transferred across datasets, then deployed a model on an NVIDIA Jetson Orin Nano.

My work: model training, cross-dataset evaluation, and Docker/TensorRT deployment.

YOLOv8m Jetson TensorRT Docker

Evaluation & deployment details

Ran 24 evaluations across a 6-model × 4-domain matrix using precision, recall, mAP50, and mAP50–95.

The CrowdHuman-only model reached 0.246 mAP50–95 on 10,000 BDD100K images. This was the strongest BDD100K result among the six training strategies tested. Transfer performance depended on the direction of the dataset shift.

For the Jetson deployment, FP16 reduced 512 × 512 inference latency from 34.4 ms to 9.6 ms while preserving 0.211 mAP50–95.

DEEP LEARNING / MEDICAL IMAGING

Mammography Classification & Segmentation

Code & evaluation results ↗

I built a pipeline for cancer classification and lesion analysis, combining ResNet-50, segmentation, and Grad-CAM with patient-level data splits.

My work: data preparation, model fine-tuning, segmentation experiments, and integrated inference.

PyTorch ResNet-50 OpenCV pydicom

Results & approach

Used 2,687 CBIS-DDSM images at 1024 × 640 resolution. Classification ROC-AUC was 0.823 on validation and 0.698 on test data.

ResNet50-U-Net transfer learning improved validation Dice from 0.153 to 0.357; test Dice was 0.369.

Evaluation lesson: the validation-to-test ROC-AUC drop (0.823 → 0.698) shows why performance must be checked on held-out patients. Segmentation experiments compare transfer learning against a baseline U-Net.

This is an independent learning and research project, not a clinical tool.

Typical held-out test segmentation cases: mammogram, ground-truth lesion mask, predicted mask, and predicted contours, with per-case Dice scores. — **Lesion segmentation: typical test cases**
Mammogram → ground-truth mask → predicted mask → predicted contours. Typical cases show both overlap and prediction errors; per-case Dice scores are included.
View full size ↗

PURDUE EPICS / EMBEDDED SYSTEMS

Geofence go-kart safety system

Built a Raspberry Pi Pico prototype integrating GPS and IMU data with a web map for defining boundaries, exporting coordinates, and monitoring vehicle position on-device.

Raspberry Pi Pico GPS / IMU Prototyping

WORK IN PROGRESS

Projects in development

Current implementation and planned next steps.

IN PROGRESS

MediaQueue

A media-processing service intended to let users upload videos, request thumbnails or compressed versions, and retrieve results while processing runs in the background.

Implemented: a Python/FastAPI API for file uploads with unique IDs, FFmpeg thumbnail generation, and JPEG retrieval.

Next: persistent jobs and background processing.

Python / FastAPI / FFmpeg

IN PROGRESS

RISC-V processor

A processor project aimed at implementing the RISC-V instruction set: fetching and decoding program instructions, performing computations, and accessing registers and memory.

Current stage: active development, early implementation. The goal is to connect computer architecture concepts to a working design and verify its behavior with test programs.

Architecture and verification details will be added as the design develops.

Computer architecture / RISC-V

03 / RESEARCH

Research experience

FEB 2026 — PRESENT

Research Assistant

Governance & Responsible AI Lab
Purdue University

Develop Python media-enrichment tools for the Political Deepfakes Incident Database, supporting research on political deepfakes and their social and policy implications.

Build modular workflows for images, video, and audio using Qwen2.5-VL, OCR, and audiovisual metadata extraction.
Map a 53-field Airtable schema into automated research workflows, including 22 model-generated fields covering visual descriptions, targets, harms, communication goals, framing, and policy narratives.

04 / EXPERIENCE

Experience

TEACHING

Supplemental Instruction Leader

Purdue Academic Success Center · Multivariable Calculus · Aug 2026–Present

Lead collaborative Multivariable Calculus sessions with 60+ students and exam reviews with 300+ students, designing activities that reinforce challenging course concepts.

TEACHING

Undergraduate Teaching Assistant

Purdue First-Year Engineering · Aug 2025–May 2026

Supported 100+ students with prototyping, MATLAB modeling, Excel analysis, and hardware troubleshooting; graded 100+ technical assignments per month.

LEADERSHIP

Treasurer

Delta Chi · Purdue Chapter · Jan–Aug 2026

Managed financial operations for a $60,000 annual chapter budget serving 50+ members.

05 / CONTACT

Contact

Reach out about internship opportunities or my projects.

sing1775@purdue.edu ↗

LinkedIn ↗ GitHub ↗ Résumé (PDF) ↓