shreshth@home — STATUS: ONLINE
Shreshth Rai
███████╗██╗ ██╗██████╗ ███████╗███████╗██╗ ██╗████████╗██╗ ██╗ ██████╗ █████╗ ██╗ ██╔════╝██║ ██║██╔══██╗██╔════╝██╔════╝██║ ██║╚══██╔══╝██║ ██║ ██╔══██╗██╔══██╗██║ ███████╗███████║██████╔╝█████╗ ███████╗███████║ ██║ ███████║ ██████╔╝███████║██║ ╚════██║██╔══██║██╔══██╗██╔══╝ ╚════██║██╔══██║ ██║ ██╔══██║ ██╔══██╗██╔══██║██║ ███████║██║ ██║██║ ██║███████╗███████║██║ ██║ ██║ ██║ ██║ ██║ ██║██║ ██║██║ ╚══════╝╚═╝ ╚═╝╚═╝ ╚═╝╚══════╝╚══════╝╚═╝ ╚═╝ ╚═╝ ╚═╝ ╚═╝ ╚═╝ ╚═╝╚═╝ ╚═╝╚═╝
AI Researcher and Applied ML Engineer
Building multimodal, agentic, and scalable AI systems across research and production.
Technical Stack
// CORE ARSENAL
Core AI & ML
Machine LearningDeep LearningGenerative AILarge Language ModelsMultimodal AIRepresentation LearningModel Fine-tuning
AI Systems & Applications
Agentic AIRAGSemantic SearchMultimodal ReasoningAI EvaluationRecommendation SystemsComputer VisionAudio AI
ML Engineering & Infra
ML InfrastructureGPU ComputingModel OptimizationDistributed SystemsInference EngineeringContainerizationMLOps
Backend & Data
PythonC/C++SQLFastAPIPostgreSQLMongoDBRedisVector Databases
Cloud & DevOps
AWSGCPDockerKubernetesNginxCI/CDSupabaseGitea
// WORKING KNOWLEDGE
Additional
ChromaDBFAISSBashJava
Selected Projects
QudaML_INFRA
Managing GPU resources across 150+ concurrent ML researchers with fair scheduling, security isolation, and minimal latency.
Built a priority-based GPU job scheduler serving 150+ users with Redis-based queuing and sub-2-second latency. Sandboxed GPU workloads using rootless Docker, GPU passthrough, and seccomp, with post-run artifact storage and presigned S3-compatible delivery.
DECIBELMM_MOE
Scaling unified audio-language models under strict compute and memory constraints while maintaining strong multimodal reasoning performance.
Designed a 7-expert Mixture-of-Experts architecture for unified audio-text reasoning under <6GB VRAM. Achieved 10.3% WER, 97% accuracy, 0.439 mAP, and +20 BLEU through parameter-efficient fine-tuning (LoRA, adapters), quantization, and distillation with Wav2Vec2 and Falcon-7B.
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 for natural-language product search and preference modeling for personalized recommendations. Implemented SPADE + DensePose virtual try-on achieving 0.77 SSIM with optimized inference pipelines for scalable retrieval.
Research Core
When Agreement Is Not Enough: A Selection Bottleneck in Non-Verifiable ReasoningECCV 2026 CDEL
Investigating failure modes in reward model agreement for non-verifiable reasoning tasks, proposing a selection bottleneck framework to improve reasoning robustness.
MedCurate-Bench: Auditing the Diagnostic Validity of Curated Medical Image DatasetsECCV 2026 CDEL
A systematic benchmark for auditing the diagnostic validity of widely-used curated medical image datasets, exposing labeling inconsistencies and curation biases.
Professional Log
Orangecat Technologies — Artificial Intelligence InternJun 2026 — Jul 2026
Architected agentic AI and RAG systems for geological and petroleum intelligence, combining multimodal semantic search, a ≤20-step LLM reasoning loop, secure sandboxed coding agents, and sub-second retrieval across 6+ document formats.
Bitrix Innovations — Full Stack AI EngineerJan 2026 — Apr 2026
Engineered and deployed a containerized AI platform on AWS EC2 using Docker, Nginx, and SSL, achieving near 100% uptime. Built LLM-driven video synthesis and multi-agent interview simulation systems, reducing API token usage by 33%.
Achievements & Awards
Winner — Delhi AI Grind
Built a multimodal civic issue platform using GraphRAG with geospatial deduplication and automated department routing.
2nd Runner-up — Hack4Delhi
Built a vision-language railway tampering detection system optimized to 20 FPS using FlashAttention 2.0 and temporal sampling.
Grand Finalist — Smart India Hackathon 2025 (DRDO)
Developed a unified Audio Language Model focused on robust, low-latency, on-premise inference.
Current Leadership
AI Head
Artificial Intelligence and Machine Learning Society (AIMS), DTU
Conduct ML knowledge-transfer sessions, teach freshmen, and mentor teams for technical hackathons.
Education
B.Tech in Mathematics and Computing
Delhi Technological University — Class of 2028
8.40 CGPA
Initiate Contact
Available for research-driven projects and AI engineering roles.