Open to Internship Opportunities

Aditya Hegde

ML Infrastructure & Applied AI Engineer

MS Artificial Intelligence student (4.0 GPA) building production-scale ML systems, GPU inference pipelines, and backend AI infrastructure. Experience spans computer vision, distributed preprocessing systems, model optimization, and geospatial AI workflows.

View Projects
Aditya Hegde — ML Infrastructure Engineer
Education

Academic Background

MS in Artificial Intelligence
San Jose State University
Aug 2025 – May 2027 GPA: 4.0
Relevant Coursework
Machine Learning Deep Learning Applied NLP AI and Data Engineering Mathematics for Data Science
B.Tech in Computer Science
Woxsen University
Sep 2021 - Apr 2025 GPA: 3.73 Dean's List (8 Semesters)
Relevant Coursework
Data Structures Web Development Artificial Intelligence Engineering Mathematics Probability and Statistics Database Management Systems
Experience

Professional Work

Infrastructure-focused ML engineering with emphasis on scale, reliability, and system design.

Data Scientist Intern

SUHORA Technologies Pvt. Ltd.
Production ML Infrastructure — NDA-safe abstraction
Mar 2024 – Jul 2024
  • Built event-driven ingestion pipeline monitoring network-attached storage for high-resolution satellite imagery (~10,000 km² per image).
  • Developed hybrid C# + Python preprocessing pipeline: orthogonalization, projection correction, normalization, uint8 simplification, and image tiling for parallel inference.
  • Implemented parallel worker architecture executing inference on dedicated GPU server with runtime memory guards and corrupted input safeguards.
  • Designed in-memory reconstruction and overlap-aware annotation merging for tiled inference outputs.
  • Generated GIS-compatible GeoJSON vector outputs for client delivery.
  • Added structured logging with timestamped exceptions, resource state capture, faulty input tracking, and automated Git-based issue logging.
  • Stress-tested on ~15 billion pixels (150 large images).
<1 min Inference per 10K km²
+40% mAP Improvement
900+ Images Processed
>95% Evaluation Accuracy

Research Assistant

Woxsen University
Aug 2023 – Feb 2024
  • Increased video frame interpolation speed by 28.5% while simultaneously improving PSNR and SSIM quality metrics.
  • Evaluated federated learning across 10+ edge devices, reducing convergence time by 15%.
  • Reduced communication overhead by 50% via distributed training optimization.
  • Built modular PyTorch + OpenCV pipelines for reusable experimentation across research projects.
  • Designed data pipelines handling 50K+ cross-domain records.
  • Authored 2 peer-reviewed publications (IEEE, Springer) on machine learning applications in space risk management and medical imaging. View publications →
Projects

Engineering Case Studies

Detailed breakdowns of production-grade systems — from problem to architecture to measurable outcome.

Marine Debris Detector

PyTorch Docker Flask React Geospatial

End-to-end satellite debris detection system processing 1000+ images. Built to identify floating ocean debris from satellite imagery at scale, with spectral index fallbacks for challenging conditions and geo-vector output compatible with industry-standard GIS tools.

>90% Validation Accuracy
+15% Precision vs Baseline
1000+ Images Processed
Dockerized Deployment
Problem & Approach

Ocean pollution monitoring requires processing large volumes of satellite imagery. Manual analysis is impractical at scale. Built a PyTorch-based detection pipeline with rule-based geospatial fallbacks (NDWI, NDVI, FDI) for spectral edge cases.

Engineering Challenges
  • Handling diverse satellite imagery conditions and resolutions
  • Balancing precision vs recall for environmental monitoring use cases
  • Integrating geo-vector output formats (QGIS/ArcGIS compatible)
  • Batch inference pipeline with REST backend serving JSON predictions
View on GitHub

Seg2Box Annotation Tool

Python JSON Processing Pip Package

Deterministic multi-stage annotation conversion pipeline for transforming segmentation masks to bounding box format. Replaced standard IoU matching with IoMin for significantly improved precision, packaged as a reusable Python module.

<1ms JSON Processing
+13% Precision (IoMin)
40% Preprocessing Reduction
Modular Reusable Package
Problem & Approach

Converting segmentation annotations to bounding box format is a common but error-prone preprocessing step. Built a deterministic pipeline with IoMin matching instead of IoU, yielding measurably better precision across diverse annotation datasets.

Engineering Challenges
  • Handling edge cases across heterogeneous annotation formats
  • Optimizing for sub-millisecond processing without sacrificing accuracy
  • Ensuring deterministic, reproducible output for CI pipelines
  • Designing clean API for packaging as reusable Python module
View on GitHub

Denoising Autoencoder

PyTorch GitHub Actions CI/CD Model Optimization

Custom lightweight CNN architecture optimized for real-time image denoising. Led a team of 12 engineers, applying pruning, quantization, and datatype optimization to achieve sub-30ms inference while maintaining high quality metrics.

35+ dB PSNR
0.95+ SSIM
<30ms Inference (512×512)
12 Team Members Led
Problem & Approach

Image denoising at inference speeds suitable for real-time applications demands carefully optimized model architecture. Designed a custom lightweight CNN with aggressive model compression techniques while preserving perceptual quality.

Engineering Challenges
  • Balancing model compression ratio vs quality metric degradation
  • Coordinating 12-person team with production-style branching workflows
  • Implementing pruning, quantization, and dtype optimization in PyTorch
  • Setting up GitHub Actions CI/CD for automated testing and benchmarking
View on GitHub
Technical Skills

Technology Stack

Core competencies across the full ML infrastructure stack.

🧠

ML / AI

PyTorch TensorFlow Transformers Scikit-Learn OpenCV HuggingFace Prompt Engineering RAG
⚙️

Backend / Systems

FastAPI Flask REST APIs SQL Batch Inference GPU Inference Structured Logging
💻

Languages

Python Java C C++ JavaScript HTML / CSS
☁️

Cloud / DevOps

Docker AWS Lambda AWS EC2 GCP Linux GitHub Actions
🔌

LLM Integrations

OpenAI API Gemini API Claude API Structured Outputs RAG Pipelines
Publications

Research Contributions

Leadership

Technical Leadership

Chief Projects Officer

AI & ML Club — San Jose State University
  • Managing 17 active projects across various ML and AI domains.
  • Leading and mentoring 8 project officers in technical delivery.
  • Standardized GitHub workflows and engineering best practices.
  • Coordinated 5 industry-led collaboration projects.
17 Active Projects
70+ Contributors
8 Officers Led
5 Industry Collabs

AI Projects Lead Officer

AI & ML Club — San Jose State University
  • Co-recruited and onboarded 6 project officers in collaboration with fellow Lead Officer and the Chief Projects Officer.
  • Established structured workflows for project planning, milestone tracking, and execution across club initiatives.
  • Developed a semester-wide project timeline to align deliverables with development milestones.
  • Coordinated cross-functional efforts with officers to organize project pitch sessions and streamline member-to-project allocation.
6 Officers Recruited
Semester Timeline Designed
Engineering Deep Dives

Under the Hood

Architecture patterns, systems thinking, and performance engineering in practice.

🏗️

Systems I've Built

Geospatial ML Inference Pipeline — Event-driven system ingesting satellite imagery, preprocessing through hybrid C#/Python pipeline, and executing parallel GPU inference. Processed 15B+ pixels at <1 min per 10K km².
Batch Image Processing Engine — Automated pipeline processing 900+ SAR & optical images with >95% evaluation accuracy, structured logging, and fault tolerance.
Annotation Conversion Pipeline — Deterministic, sub-millisecond processing for converting between annotation formats used in object detection workflows.
🔧

ML Infra Patterns

Event-Driven Ingestion — File system monitoring triggers for automated pipeline initiation on new data arrival.
Parallel Worker Architecture — Distributing inference workloads across GPU resources with load-aware scheduling.
Runtime Memory Guards — Preventing OOM failures through proactive resource monitoring and detection overflow limits.
Structured Logging — Timestamped exceptions, resource state capture, and automated Git-based issue tracking.
⚡

Performance Notes

Inference Latency — ~7 min → <1 min (85%+ reduction) through pipeline architecture redesign.
Model Accuracy — +40% mAP via systematic model selection and data-centric iteration.
Distributed Training — 50% communication overhead reduction and 15% faster convergence.
Model Compression — Sub-30ms inference on 512×512 via pruning, quantization, and dtype optimization.
Get in Touch

Let's Connect

Open to ML infrastructure, applied ML, and backend engineering internship opportunities. Reach out to discuss systems, research, or collaboration.

aditya.hegde@sjsu.edu GitHub LinkedIn