Shreyansh Jain
01
Education
SRM Institute of Science and Technology
Bachelor of Technology in Computer Science and Engineering, CGPA: 9.54, Merit Scholarship(2×)
2023 -- 2027
Chennai, India
02
Research & Industry Experience
AI Research Intern
O6AI Labs Pvt Ltd
Jul 2025 -- Present
Remote
- Conducted research on fine-tuning a 500M-parameter language model using LoRA and GRPO, designing custom reward functions and systematic evaluation protocols; resulting work published in an IEEE conference.
- Designed and developed O6-Suite, an agent-driven command center built around a ``less UI, more intelligence'' paradigm, unifying data and event flows through a shared computational substrate; engineered a graph-based state layer for evolving entities and relationships and a mathematically-driven B2B ranking system based on relevance, urgency, and user preferences.
- Developed autonomous multi-agent pipelines for Beacon, O6's recruitment vertical, automating workflows from screening to voice agents with minimal human-in-the-loop; applied DSPy, GEPA, CodeAct, and RLM-inspired architectures to reduce token usage by 40% while improving long-context reasoning reliability.
- Contributed to several additional internal AI systems and product projects across O6's verticals, spanning research, architecture, and development.
Computer Vision Engineer Intern
Siemas Technology Pvt Ltd
Aug 2024 -- Nov 2024
Chennai, India
- Implemented a real-time facial recognition system achieving 30+ FPS and sub-1ms latency, supporting 2+ concurrent streams through algorithmic optimization, multithreading, frame serialization, and ZeroMQ.
03
Selected Research & Projects
OmniLock: Creator-Owned Provenance System
C++, CUDA, Rust, Python, IPFS- Researching a computational provenance system for video that embeds a 64-bit signed identifier into motion-compensated residuals using mid-frequency DCT bands, LDPC error correction, and deterministic cryptographic verification.
- Engineering a high-performance C++/CUDA implementation with GPU-accelerated signal processing and zero-copy memory pathways, combining active watermark extraction with passive perceptual fingerprinting to resolve media to creator-owned license records.
FP-SLM: Physics-Informed Small Language Model
PyTorch, CUDA, PDEs, PINO, COCONUT- Designing a hybrid scientific machine learning architecture coupling a pretrained language model with a latent reasoning engine to translate textual descriptions of physical systems into solutions of partial differential equations.
- Developing a computational pipeline combining latent representation learning with a Physics-Informed Neural Operator, enforcing PDE residual constraints during training and leveraging GPU acceleration for iterative latent reasoning and operator evaluation.
LT-GRPO: Latent Thought Group Relative Policy Optimization
PyTorch, Transformers, CUDA, JEPA, GRPO- Developed a latent reasoning architecture that intercepts intermediate LLM representations and maps them into a 512-dimensional continuous thought space, using a Gaussian policy, JEPA-based transition model, adaptive computation halting, and group-relative policy optimization to reduce reliance on token-level chain-of-thought.
- Extended the framework into an ongoing study of layer-wise reasoning emergence, developing mathematical and empirical methods to identify reasoning-relevant transformer layers rather than assuming a fixed middle-layer hypothesis; initial experiments show improvement over the baseline architecture.
Context-Aware RAG System
Python, LangChain, ChromaDB- Built a metadata-enhanced RAG pipeline inspired by Anthropic's Contextual Retrieval research, implementing custom context generation for document chunks.
- Achieved 35% reduction in retrieval failure rate through dynamic metadata filtering and semantic scoring algorithms.
04
Publications
Mathematical Framework for Custom Reward Functions in Job Application Evaluation using RL
- Developed an SFT + GRPO training pipeline for a ~502M-parameter language model, achieving 91.4% accuracy through custom multi-component reward design.
Quantum-Enhanced Smart Computing Framework for Sustainable Credit Risk Decision
- Developed a hybrid quantum-classical framework achieving 94.53% accuracy on 32,000+ loan records.
Hybrid Quantum Model for Digital Media Processing for Metal Surface Defect
- Integrated quantum variational circuits with CNNs using PennyLane, achieving 99.94% accuracy.
05
Selected Achievements
Smart India Hackathon Finalist
ISRO Problem Statement
2025
- Selected as finalist among 800,000+ participants; developed an AI/ML forecasting model for ground-level O₃ and NO₂ achieving RMSE 8.4 and R² = 0.92.
- Technical Head, Innovation Incubation & Entrepreneurship Cell, SRM --- led 9+ technical projects and organized hackathons with 200+ participants.
06
Technical Skills
Languages: Python, C++
AI/ML: Machine Learning, Deep Learning, NLP, Computer Vision, Reinforcement Learning, Scientific ML
Frameworks: PyTorch, TensorFlow, Keras, Hugging Face, LangGraph, FastAPI
Methods: LoRA, QLoRA, GRPO, RAG, JEPA, PINO, Multi-Agent Systems, Graph-based Modeling
Systems: GCP, Docker, Git, CUDA, WebSockets
07
Certifications & Expanded Coursework
Stanford CME 295: Transformers & Large Language Models --- Afshine & Shervine Amidi
MIT OCW 18.06: Linear Algebra --- Prof. Gilbert Strang
Deep Learning Specialization: DeepLearning.AI
Machine Learning Specialization: DeepLearning.AI
SHREYANSH JAIN2026