Aryan Rastogi

CPP-2

Hewlett Packard Enterprise • India

Shortlisted for the CPP-2 cohort and collaborating on an enterprise project with HPE.

Amazon ML Summer School

Amazon • 2026

Selected for the prestigious Amazon ML Summer School 2026 program, an intensive machine learning education initiative by Amazon. One of 3,000 students selected from over 1.4 lakh applicants nationwide.

AI Engineer

Sylus AI • (remote)

• Engineered the platform end-to-end from a greenfield repository to production, owning auth, database schema, and core architecture through 180+ commits, then scaling it into a multi-tenant AI content platform.
• Designed and shipped a 9-phase agency system (RBAC, approval workflows, client portal, invites, real-time collaboration) from scratch, taking the product from single-user to multi-brand agency-ready.
• Built the AI content engine (multi-model LLM copywriting, live Google News trend discovery, AI image generation) and a document intelligence pipeline auto-extracting and scheduling posts from PDFs, Excel, and Word docs.
• Hardened the platform for production: patched security vulnerabilities across auth, API endpoints, and middleware (SSRF, CORS), and stabilized deployment across Cloud Run and Docker.

SDE Intern

Southern Federal University, ICTIS • Russia (remote)

Developed a web application as an international interdisciplinary project in collaboration with Southern Federal University (ICTIS) & MUJ to support the adaptation of international students. The platform acts as a digital assistant that helps to explore, understand and navigate the campus facilities and improves accessibility for new international students.

AI R&D Intern

Elumina Street Solutions • Jaipur

Developed a violence detection model using CNN (MobileNet) and LSTM in TensorFlow for video-based behavior recognition; improved accuracy with focal loss and class weighting. Implemented and evaluated a real-time weapon detection system using YOLOv8 and OpenCV. Collaborated with the AI research team to analyze model performance, optimize hyperparameters, and enhance detection.

Research Intern

AI²: computAtional linguIstics & vIsion Lab • MUJ

• SafeAlign: Leading a multi-arm experimental pipeline for LLM safety evaluation using taxonomy-guided data curation on the Anthropic HH-RLHF dataset, training a RoBERTa-based reward model, a QLoRA+DPO-aligned Qwen2.5-1.5B generative model, and a MiniLM contrastive safety encoder.
• Key Finding: Discovered that a 66M DistilBERT classifier trained on taxonomy-filtered data achieves 81.3% accuracy vs 59.5% on raw data, proving label quality outweighs model scale.
• Quantifying the Alignment Tax: Empirically measuring how SFT and DPO degrade LLM creativity using DAT, AUT, and TTCT benchmarks with Semantic Entropy metrics across Mistral-7B and Llama-3.1-8B.