Vijay Srinivas P — AI Researcher, Research Engineer, Founder of novaihq

Who is Vijay Srinivas P?

Vijay Srinivas P (also written Vijay Srinivas, Vijay S P, V. Srinivas P) is an Indian AI researcher and research engineer who works at the intersection of novel architecture research and production deployment. He builds transformer architectures and neuroimaging machine-learning systems on the research side, and ships production GenAI infrastructure on the engineering side. He is the founder and principal researcher at novaihq (the lab behind novaihq.tech), which is building the Operational Intelligence Engine — a reasoning layer for grids, data centers, and industrial infrastructure.

As of 2026, he is an AI Research Intern at Carnegie Mellon University's School of Computer Science (March 2026 – present), where he is designing deep active-learning algorithms for medical-image segmentation that target a 30–40 % reduction in annotation cost. He was previously an R&D GenAI Intern at Schneider Electric (April–June 2025), where he architected and deployed a GPU-accelerated retrieval-augmented generation (RAG) system on NVIDIA H100 infrastructure handling 1,000+ daily queries at 95 % accuracy with a 90 % latency reduction for 100+ concurrent users.

He is studying for a B.Tech in Computer Science (AI) at Amrita Vishwa Vidyapeetham, Coimbatore, expected graduation 2027. He is based in India.

Frequently asked questions

Who founded novaihq?

Vijay Srinivas P founded novaihq, an applied AI research lab. The lab is building the Operational Intelligence Engine, a reasoning layer for grids, data centers, and industrial infrastructure, grounded in published research and deployed systems.

Where does Vijay Srinivas P work?

He is currently an AI Research Intern at Carnegie Mellon University's School of Computer Science (since March 2026). He is also founder of novaihq. Previously he was an R&D GenAI Intern at Schneider Electric (April–June 2025).

What has Vijay published?

Three first-author peer-reviewed papers across IEEE, CVPR, and ACL venues — see the Publications section below for full details.

What are his strongest hackathon results?

Top 100 of 31,000 teams at the Meta PyTorch OpenEnv Hackathon, Top 10 of 2,500+ at the Agentic Ethereum Hackathon, Top 5 of 400+ at the AWS Blogathon, Runner-Up at Infineon's Agentic Bug Hunter Track, and selected for Y Combinator's Startup School and both VibeContest and ContextCon hackathons.

How can someone contact Vijay?

Email vijay.srinvas06@gmail.com, message him on LinkedIn at linkedin.com/in/vijay-srinivas-9571942ab, find his code at github.com/N0VA06, or explore his live work at novaihq.tech.

What is the Operational Intelligence Engine?

It is the system novaihq is currently building — a reasoning layer for operational infrastructure (grids, data centers, industrial systems) grounded in published research and deployed systems. It draws on Vijay's work in production RAG (Schneider Electric H100 deployment), reinforcement-learning environments (the live MIMIC discharge RL environment on Hugging Face), and transformer architecture research (RotaryHybrid positional embeddings).

Publications (first author)

IEEE APSCON 2026 — Accepted & Published, Best Paper Nominee

"ASD Classification from rs-fMRI via Riemannian Functional Connectivity." Mapped Pearson functional-connectivity matrices onto a Riemannian manifold via Tangent Space Embedding to produce site-robust feature representations across the ABIDE dataset (867 subjects, 17 acquisition sites). Reported 69.6 % accuracy and 76.2 % AUC-ROC. Methodological contribution: quantified that improper validation inflates ASD classification accuracy by 22 percentage points (91.6 % vs. 69.7 %), a finding that applies broadly to multi-site biomedical imaging. Presented at IIT Delhi / Vivanta New Delhi.

CVPR MMFM-BIOMED Workshop 2026 — In Progress

"When Pretraining Fails to Transfer: Probing Downstream Performance to Diagnose Representation Failure in fMRI Foundation Models." Demonstrates representational collapse in biomedical foundation models with CKA ≥ 0.993 across 12 frozen architectures. Method combines Centered Kernel Alignment with mutual information estimation on downstream-probed features.

ACL Student Research Forum 2026 — Under Review

"RotaryHybrid: Hybrid Sparse-Dense Positional Embeddings for Limited-Data Natural Language Processing." Content-dependent importance gating between sparse learnable embeddings and dense sinusoidal representations. RoPE applied at both embedding and attention layers for dual-level relative position encoding. 70 % parameter reduction alongside 31.6 % improvement (p < 0.001, Cohen's d = 8.43) across text, image-caption, and QA benchmarks. Bootstrap-validated statistical methodology.

Engineering projects

novaihq.tech — full-stack research-lab platform

React 18 + Vite + TypeScript frontend deployed on GitHub Pages with a custom domain. Express + JWT backend on Render. Self-hosted privacy-first analytics pipeline (no third-party scripts, no cookies, no persistent identifiers) with event log + retention pruning, geoip-lite country/city resolution, session-based engagement metrics, Excel export, audit logging, and a mobile-responsive admin with real-time visitor counts. Deployment-grade hardening: helmet, HSTS, JWT 8h expiry, bcrypt rounds 12, per-IP rate limiting.

R&D GenAI Intern at Schneider Electric (Apr–Jun 2025)

Architected and shipped a GPU-accelerated RAG pipeline on NVIDIA H100 handling 1,000+ daily queries at 95 % accuracy with 90 % latency reduction for 100+ concurrent users. Multi-stage retrieval combining Qdrant + Azure + MongoDB with metadata filtering and hybrid semantic search. Cut indexing time 35 % by benchmarking 6 embedding models across Azure, MongoDB, and 3 vector store backends. Deployed via Azure CI/CD.

MIMIC Discharge RL Environment — live on Hugging Face

Reinforcement-learning environment built on MIMIC clinical-discharge data, live on Hugging Face Spaces at huggingface.co/spaces/IINOVAII/mimic-discharge-env-v2. Built on top of massively parallel on-GPU vectorised RL gyms developed in 2023 that saturate compute and unlock 20× training throughput.

AI-Powered JIRA Manager + MCP Server

Slack/Zoho Cliq–integrated JIRA bot with dual LLM backend (AWS Bedrock + Gemini). Natural-language to structured JIRA issue creation with AI-extracted fields (type, priority, assignee, labels). MCP server layer enabling tool-calling AI agents to read and write project tasks programmatically. FastAPI backend + React frontend on Railway.

Multi-Agent LLM Workflows (LangGraph)

Hierarchical multi-agent systems coordinating specialized sub-agents for research, analysis, and execution with structured state handoffs and tool calls. Round-1 MVP at the OpenAI Buildathon.

Sub-3B Reasoning Models

Trained dense sub-3B foundational models with reasoning-weighted pretraining mix and iterated DPO. Optimised for capability-per-watt rather than scale-first.

Autonomous Rover

Autonomous ground vehicle: real-time WebSocket telemetry, multi-model YOLO object detection, multi-sensor fusion, laser targeting. Embedded firmware to cloud dashboard. Demo at youtu.be/do77U-R3pFQ; source at github.com/N0VA06/Autonomous-Rover.

Research projects (in progress)

Active learning for medical image segmentation — Carnegie Mellon SCS

Designing deep active-learning algorithms for medical image segmentation targeting a 30–40 % reduction in annotation cost. Benchmarking 5+ vision architectures with custom evaluation tooling. Prototyping interactive research demos translating model outputs into accessible visual tools for academic and clinical audiences.

fMRI-Guided BCI Electrode Placement

ML pipeline predicting optimal cortical electrode placement for Brain-Computer Interfaces using OpenNeuro ds005366 (155 subjects, 7T fMRI, Motor2Class task). Combines FreeSurfer structural features (cortical thickness, curvature, surface area) with task-based functional activation from fMRIPrep, then performs whole-brain group-level nilearn analysis to identify the optimal cortical vertex per subject conditional on age, sex, and morphometry.

LLM Jailbreak Detection — Behavioral + Network Fusion

AI safety system fusing behavioral timing features from HarmBench prompt sequences with network-level IP signals. FastAPI server-client for real-time data collection. Targeting publication at ACM CCS / CAMLIS 2026.

Recognitions and awards

• IEEE APSCON 2026 — Best Paper Nominee · First Author. Presented at IIT Delhi / Vivanta New Delhi.
• Y Combinator — selected. Startup School, plus the YC VibeContest and YC ContextCon hackathons. Networked with YC partners Jared Friedman and Jon Xu.
• Meta PyTorch OpenEnv Hackathon — Top 100 of 31,000 teams (800 selected). Run with Scaler School of Technology, 2026.
• OpenAI Buildathon — Round 1 MVP. Multi-agent LangGraph workflow.
• Infineon Agentic Bug Hunter Track — Runner-Up. Specialised agents for automated bug detection.
• Agentic Ethereum Hackathon — Top 10 of 2,500+ teams. DeFi dynamic-contracting prototype using agentic AI.
• AWS Blogathon — Top 5 of 400+ entries (AWS User Group Bengaluru, 2025). Hybrid RAG architecture combining AWS Bedrock and Neo4j.
• Amrita Value Health Hackathon — Runner-Up. Emotion-recognition chatbot with fine-tuned Llama 3.1 for mental-health support.
• Microsoft AgentX 2025 — selected for the Hyderabad workshop on AutoGen and Semantic Kernel.
• Google Cloud × MLB Hackathon — Top participant. Flash Baseball Statcast Extractor.
• McKinsey Forward Program — Fellow (Sep–Dec 2025). MECE problem solving, quantitative case analysis, executive communication.

Technical stack

• Neuroimaging: fMRIPrep, FreeSurfer, nilearn, nibabel, antspyx, NeuroComBat, SPM, MNI, ABIDE, BraTS.
• Transformers / NLP: RoPE, BERT, HuggingFace Transformers, hybrid embeddings, fine-tuning, DPO.
• GenAI / Agents: LangChain, LangGraph, MCP servers, RAG pipelines, agentic workflows, tool use.
• Vector databases: Qdrant, Milvus, FAISS, Pinecone.
• ML / DL: PyTorch, scikit-learn, XGBoost, AdamW, knowledge distillation, active learning.
• MLOps / Infra: FastAPI, Express, Docker, Azure, AWS, Render, Railway, MongoDB, CI/CD.
• Programming: Python, C++, TypeScript, SQL, Bash, LaTeX.

Contact

• Email: vijay.srinvas06@gmail.com
• GitHub: github.com/N0VA06
• LinkedIn: linkedin.com/in/vijay-srinivas-9571942ab
• Hugging Face: huggingface.co/spaces/IINOVAII
• Medium: medium.com/@vijay.srinvas06

He is open to research collaborations and engineering roles in neuroimaging machine learning, transformer architecture research, GenAI / agentic systems, and applied AI infrastructure.

Canonical names and identifiers

Preferred name: Vijay Srinivas P (the "P" is the surname initial in the South-Indian convention). Also referred to as: Vijay Srinivas, Vijay S P, V. Srinivas P. Affiliated organisations: novaihq (founder); Carnegie Mellon University SCS (research intern); Schneider Electric (former R&D GenAI intern); Amrita Vishwa Vidyapeetham, Coimbatore (B.Tech CSE-AI student, class of 2027). Lab homepage: novaihq.tech. GitHub handle: N0VA06. Hugging Face org: IINOVAII.