Madhav S. Baidya

AI Researcher | LLM Multi Agentic Systems | Multimodal Reasoning | Knowledge Graphs

---

About

Final-year undergraduate at IIT (BHU), Varanasi working at the intersection of LLMs, multimodal reasoning, and knowledge-grounded systems. My work focuses on building epistemically reliable AI systems, combining structured representations (knowledge graphs) with retrieval and generation pipelines.

I am particularly interested in:

• Decision-aware retrieval and generation (Decision RAG)
• Multi-agent LLM systems for reasoning and control
• Multimodal information extraction and grounding
• Robustness, evaluation, and failure analysis of LLM systems

---

Research Interests

• Large Language Models (LLMs) and Agentic Systems
• Multimodal Learning (Text, Vision, Audio)
• Knowledge Graphs and Neuro-Symbolic AI
• Retrieval-Augmented Generation (RAG) and Decision RAG
• AI Robustness, Hallucination Mitigation, and Evaluation

---

Research & Publications

• PassiveQA: A Three-Action Framework for Epistemically Calibrated Question Answering via Supervised Finetuning
  

  • Introduced a planner-driven multi-agent system with ASK / ANSWER / ABSTAIN routing
  • Designed a decision RAG mechanism over a knowledge graph with query-guided edge weighting
  • Incorporated explicit variable injection (?var) to model missing information and enable structured multi-hop reasoning
  • Fine-tuned a Mistral-7B planner (LoRA) on a graph-grounded dataset for improved abstention and reduced hallucination

• Detecting the Machine: A Comprehensive Benchmark of AI-Generated Text Detectors Across Architectures, Domains, and Adversarial Conditions
  

  • Built a large-scale evaluation framework across transformers, CNNs, stylometric models, and LLM-based detectors
  • Evaluated under domain shift, cross-LLM generalization, and adversarial humanization
  • Identified systemic limitations such as detector–generator coupling and robustness failure

Selected Work

• Multimodal Knowledge Graph System
  RDF-based graph construction (RDFlib + SPARQL) from unstructured documents with hybrid retrieval (symbolic + vector) for multi-hop QA

• Agentic Research Assistant (LangGraph)
  Multi-agent pipeline for automated paper retrieval, parsing, and synthesis across scientific sources

• SPAWN: Spoken Environment World Modeling
  Benchmark framework for evaluating spatial reasoning in multimodal LLMs from spoken/textual inputs, including tasks such as map reconstruction, relational inference, and navigation planning under noisy and ambiguous conditions

• SQLPilot (NL → SQL Compiler)
  Schema-aware query generation with validation and execution across MySQL + SQLite backends

• Multimodal PDF RAG System
  CLIP-based image embeddings + LLM summarization for joint text-image retrieval

---

Technical Stack

AI / ML PyTorch • TensorFlow • Representation Learning • Multimodal Learning

LLMs & Agents LangChain • LangGraph • OpenAI • Mistral • Tool-Augmented Agents

Retrieval & Knowledge Systems Qdrant • Knowledge Graphs (RDF, SPARQL, RDFlib) • Sentence Transformers

Systems FastAPI • Docker • Async Systems • Microservices • DVC

---

Activity

---

Links

• GitHub: https://github.com/MadsDoodle
• HuggingFace: https://huggingface.co/Moodlerz
• Portfolio: https://aidoodler.vercel.app
• Email: madhavbaidyaiitbhu@gmail.com

---

Note

Focused on building systems that know when they know, and when they do not.