AI researcher and research engineer building the Operational Intelligence Engine at novaihq. Currently AI Research Intern at Carnegie Mellon University SCS. First author at IEEE APSCON 2026 (Best Paper Nominee), CVPR MMFM-BIOMED 2026, and ACL 2026.
Vijay Srinivas P works at the boundary between novel architecture research and production deployment. On the research side: transformer architecture design (positional embeddings, hybrid sparse-dense representations), neuroimaging ML (fMRI, BraTS, multi-site harmonization), and biomedical computer vision. On the engineering side: production RAG systems, multi-agent LLM workflows, and MCP server integrations.
Studying B.Tech in Computer Science (AI) at Amrita Vishwa Vidyapeetham, Coimbatore — Class of 2027. Previously R&D GenAI Intern at Schneider Electric (Apr–Jun 2025), where he shipped a GPU-accelerated RAG pipeline on NVIDIA H100 handling 1,000+ daily queries at 95% accuracy.
Tangent Space Embedding on the ABIDE dataset (867 subjects, 17 acquisition sites). 69.6% accuracy / 76.2% AUC-ROC. Methodological contribution: improper validation inflates accuracy by 22 percentage points — applies broadly across multi-site biomedical imaging.
Diagnosing representation failure in fMRI foundation models via downstream probing. CKA ≥ 0.993 across 12 frozen architectures — exposing representational collapse in biomedical foundation models.
Content-gated hybrid of sparse + sinusoidal embeddings with dual-level RoPE. 70% parameter reduction alongside 31.6% improvement (p < 0.001, Cohen's d = 8.43).
React + Vite + TypeScript on GitHub Pages. Express + JWT backend on Render. Self-hosted privacy-first analytics, Excel export, mobile-responsive admin, audit logging, bcrypt-12, HSTS, per-IP rate limiting.
Multi-stage retrieval with Qdrant + Azure + MongoDB. 1,000+ daily queries · 95% accuracy · 90% latency reduction · 35% indexing speedup via benchmarking 6 embedding models across 3 vector stores. Apr–Jun 2025.
Reinforcement-learning environment for clinical-discharge decision-making, live on Hugging Face Spaces. Built on top of massively parallel on-GPU vectorised RL gyms.
Slack/Cliq-integrated bot with dual-LLM backend (AWS Bedrock + Gemini). MCP server layer so tool-calling agents can read/write project tasks. FastAPI + React on Railway.
Hierarchical multi-agent systems with structured state handoffs and tool calls. Round-1 MVP at the OpenAI Buildathon.
This page (/about) is the canonical machine-readable profile, served as static HTML
with embedded schema.org/Person JSON-LD. A short overview is also at
/llms.txt following the
llmstxt.org standard.