Vedant Sanjay Jadhav
Specializing in Mamba-2 & Transformer LLMs, Agentic AI workflows, Multimodal systems, and high-performance RAG & MLOps.
Where research
meets production
I am an AI Engineer specializing in Generative AI, Large Language Models (LLMs), Vision Language Models (VLMs), Domain-Specific Language Models (DSLMs), Agentic AI, and Physical AI.
My work spans building enterprise-grade RAG systems, multi-agent architectures, state-space models (Mamba-2), and high-performance inference pipelines with PyTorch, LangGraph, FastAPI, Qdrant, Docker, and NVIDIA GPUs.
As a published researcher, my contributions focus on state-space models, multi-agent systems, healthcare AI alignment, and embodied intelligence.
Generative & Agentic AI
01Architecting multi-agent TaskGraph workflows, autonomous agents, and hybrid RAG systems using PyTorch, LangGraph, FastAPI, and Qdrant.
Foundation & Vision Models
02Fine-tuning and deploying LLMs, VLMs (Vision Language Models), and DSLMs (Domain-Specific Language Models) for production workloads.
State Space Models & Physical AI
03Pioneering Mamba-2 SSM architectures, physical planning controllers, robotics action prediction, and ultra-efficient edge AI inference.
Featured projects
A selection of AI systems, language models, and autonomous frameworks built from concept to deployment.
SparkAI-50M-A100
Hybrid Mamba-Transformer LLM
Designed and trained a 50M-parameter LM from scratch using a hybrid Mamba-2 + GQA architecture on a single NVIDIA A100 80GB GPU, implementing tokenization, training, checkpointing, and evaluation.
- •Pretrained 50M-parameter language model from scratch
- •Hybrid Mamba-2 state-space model + Grouped Query Attention (GQA)
- •Single NVIDIA A100 80GB GPU execution
Vidya Educational LLM
Multilingual Educational AI
Developed a 1.7B-parameter multilingual Educational LLM fine-tuned for NCERT Science and Mathematics across 11 Indian languages, achieving 93.3% aggregate accuracy on a custom 64-question, 8-script evaluation suite.
- •1.7B parameters fine-tuned for NCERT Science & Math
- •Supports 11 Indian languages across 8 scripts
- •93.3% aggregate accuracy on custom evaluation suite
WinFix OmniAgent
Agentic Windows Diagnostics
Built an evidence-driven agentic troubleshooting system with planner–specialist TaskGraph execution, domain-specific diagnostic agents, safety-gated tool execution, and a real-time Gradio control console.
- •Planner–Specialist TaskGraph multi-agent execution
- •Safety-gated system tool execution
- •Real-time Gradio diagnostic control console
AutoStack AI
Autonomous ML Lifecycle Platform
Built a production-first AutoML platform that transforms raw CSV data into optimized models and low-latency REST APIs in under 60 seconds, automating profiling, feature engineering, tuning, and deployment.
- •Transforms raw CSV into REST APIs in under 60 seconds
- •Parallel optimization over Gradient Boosted Trees & Stacking Regressors
- •SHAP-based explainability, sub-10ms inference, drift monitoring
TinyVLM
Lightweight Vision Language Model
Built a lightweight VLM integrating a vision encoder with a compact LM for image captioning and visual question answering, implementing feature projection, cross-modal alignment, and instruction-following generation.
- •Vision encoder integrated with compact LM
- •Cross-modal alignment & feature projection
- •Instruction-following image captioning & VQA
Academic contributions
Ensemble and Hybrid ML Approaches for Renewable Energy Forecasting
Constructed an ensemble ML forecasting framework linking real-time model outputs with power grid stability. Reduced forecasting errors by ~30% in trials and received the Best Paper Award at ICCTVB-25.
PhysicalMamba: State Space Models for Physical Planning and Robotic Task Execution
Applied state-space models (Mamba-2) to physical trajectory planning and action prediction under hardware compute constraints, establishing linear scaling over long-horizon trajectories.
RecursiveMAS: A Recursive Multi-Agent Latent Coordination Framework for Embodied Task Optimization
Architected a Latent Coordination Framework using semantic recursion loops, enabling multi-agent synchronization and reducing communication bottlenecks across complex embodied agent tasks.
Small Language Models for Clinical Reasoning and Medical Decision Alignment
Evaluated clinical reasoning in Small Language Models (LLaMA, BioMistral, Med42v2, Qwen), defining the Semantic Drift Score (SDS) and Structural Alignment Index (SAI) to prevent clinical hallucinations.
Work experience
Tech Mahindra• AI Engineer Intern
- →Led the AI team as a Young AI Engineer on the development of Indus, contributing to model experimentation, evaluation, AI system design, and technical implementation.
- →Worked on DSLMs, VLMs, and Generative AI systems, building pipelines spanning data prep, retrieval, inference, evaluation, and deployment; worked with NVIDIA Nemotron 3.5 for LLM experimentation under practical compute constraints.
- →Designed and implemented RAG, semantic retrieval, and agentic AI workflows using Python, PyTorch, Hugging Face, LangGraph, FastAPI, and Ollama.
DPulseAI Pvt. Ltd.• AI Engineer Intern
- →Built production Generative AI and RAG systems, improving retrieval accuracy by 35% and reducing latency by 45% through hybrid chunking and reranking strategies.
- →Architected Dockerized LLM inference pipelines with CI/CD and canary deployments, cutting deployment time to under 10 minutes using GitHub Actions and Docker Registry.
- →Implemented KV-cache quantization and post-deployment drift monitoring, reducing inference memory requirements and improving LLM serving efficiency on consumer-grade hardware.
Pimpri Chinchwad University• B.Tech. in AI & Machine Learning
- →Specializing in Artificial Intelligence and Machine Learning with a aggregate CGPA of 8.10 / 10.0.
- →Active researcher in State Space Models (Mamba-2), Agentic AI, and Small Language Models (SLMs).
Skills & technologies
LLM & Generative AI01
Model Architectures02
AI Frameworks03
AI Systems & MLOps04
Data & Retrieval05
Languages06
Hackathons & achievements
Code4Society Hackathon 2026
National Level Hackathon Winner
CodeApex 24-hr Hackathon
VIT Pune National Hackathon
ICCTVB-25 Scopus Conference
Sanjay Ghodawat University
National DevCraft Hackathon
IIT Indore (Fluxus Fest)
National AI Hackathon
IIT Indore AI Track
Visual playground
Moments from hackathons, research conferences, and engineering showcases.
.jpeg&w=3840&q=75)
.jpeg&w=3840&q=75)
.jpeg&w=3840&q=75)
.jpeg&w=3840&q=75)
