shreshth@home — STATUS: ONLINE

Shreshth Rai

AI Researcher and Applied ML Engineer

AI Researcher and ML Engineer building multimodal systems and robust model architectures for real-world deployment.

Technical Stack

// CORE ARSENAL
Core ML / AI
PyTorchTensorFlowTransformersPEFTQuantizationDistillation
Multimodal & CV
AEs/VAEsGANsVision TransformersDiffusion ModelsSelf-Supervised LearningVision Language Models
Backend / Systems
FastAPINode.jsPostgreSQLMongoDBRedis
Programming
PythonCC++JavaScriptSQLBash
Data Science
NumPyPandasScikit-learnMatplotlibSeabornFAISS
Frontend
HTMLCSSReactNext.jsTypeScript
Cloud/Infra
AWSGCPAzureDockerKubernetesNginx
// WORKING KNOWLEDGE
Miscellaneous
ChromaDBJavaKubernetesKotlin

Selected Projects

DECIBELMM_MOE
Scaling unified audio-language models under strict compute and memory constraints while maintaining strong multimodal reasoning performance.
Designed a Mixture-of-Experts multimodal architecture with 7 specialized expert models. Applied parameter-efficient fine-tuning (LoRA, adapters), quantization, and distillation to reduce VRAM footprint (<6GB) while preserving performance. Integrated Wav2Vec2 embeddings and fine-tuned Falcon-7B for joint audio-text reasoning.
PyTorchTransformersLoRAPEFTQuantizationDistillationWav2Vec2MoEVIEW →
FashionateMM_GAN
Learning high-fidelity multimodal representations for personalized fashion recommendation and realistic virtual try-on.
Built a multimodal recommendation system using vision-language embeddings and large-scale similarity search. Implemented GAN-based virtual try-on with SPADE and DensePose for garment warping. Optimized inference pipelines and embedding efficiency for scalable retrieval and personalization.
PyTorchCLIPGANsDensePoseFAISSComputer VisionVIEW →
Vision-SSLSSL_CV
Reducing dependence on labeled data for high-performance visual representation learning.
Implemented self-supervised learning pipelines using SimCLR and Masked Autoencoders (MAE) on ImageNet-100. Studied representation quality under limited supervision and optimized training efficiency through contrastive learning and masked modeling approaches.
PyTorchSimCLRMAESelf-Supervised LearningComputer VisionVIEW →

Research Core

Inference Time Expert Subsets for Gradient Misaligned Robustness (Under Review)
Research in adversarial machine learning focused on improving robustness of deep models under worst-case attacks using ensembling and inference-time methods.

Professional Log

Orangecat TechnologiesArtificial Intelligence InternJun 2026 — Present
Bitrix InnovationsFull Stack AI EngineerJan 2026 — Apr 2026
Engineering LLM-driven video synthesis pipelines. Optimized AWS deployment workflows resulting in 40% latency reduction. Leading technical integration of generative models into production environments.
Department of Applied Mathematics, DTUResearch InternMay 2025 — Jun 2025
Pioneered multimodal sentiment analysis using BiLSTM and Vision Transformers (ViTs). Focused on novel feature extraction from unstructured audio and visual data streams.
5G Research Lab, DTUResearch InternJun 2025 — Jul 2025
Built AI-driven applications using 5G-enabled vision systems. Developed facial recognition pipelines with Siamese Networks. Studied 5G SA core architecture and its integration with real-time edge inference and SDR systems.

Initiate Contact

Available for research-driven projects and AI engineering roles.