Knowledge-grounded AI agents
Designed a manager–planner–executor system that uses an ontology and graph of APIs to guide retrieval, plan tool sequences, and produce traceable multi-hop decisions.
I'm Vritansh Kamal.
I am an AI researcher and data scientist working on deep reinforcement learning, post-training, neuro-symbolic AI, and trustworthy machine learning. I build systems that reason over structured domain knowledge, use tools reliably, and make their decisions easier to inspect.
I am from Kullu, Himachal Pradesh. I studied Computer Science and Engineering at SRM Institute of Science and Technology and earned an M.S. in Data Science from Columbia University. My work connects research with real systems in healthcare, telecom, and enterprise AI.
Research interests·Research projects·Teaching·Technology stack·GitHub·LinkedIn
VK
Photo to be added
Affiliations
Research interests
- Deep reinforcement learning for decision-making and learning under constraints
- Post-training methods for language models, including efficient adaptation and reward-guided learning
- Neuro-symbolic AI, knowledge graphs, and multi-hop reasoning
- Bias, fairness, and interpretability in machine-learning models
- Medical AI for high-risk scenarios, including uncertainty quantification and reliable decision support
- Applied mathematics for optimization, graph algorithms, and machine-learning systems
Research projects
Post-training small language models for structured reasoning
Exploring parameter-efficient adaptation and group-relative policy optimization for choosing the next node in a graph-constrained reasoning path.
Master’s research: Transformers for medical imaging
Researched transformer-based detection methods for PET/CT lesion analysis, comparing Conditional DETR with YOLO-style approaches for more efficient convergence and detection performance.
Large-scale graph visualization and evaluation
Designed graph-visualization and evaluation workflows for very large graphs, using force-directed layouts such as Fruchterman–Reingold alongside clustering and graph analytics to make complex relationships usable at scale.
HIPPO
Contributed as a founding member and research assistant, supporting collaborative research and applied AI exploration.
Teaching experience
Teaching Assistant II, Quantitative Methods in the Social Sciences (QMSS)
Appointed as a Teaching Assistant II for Columbia’s Quantitative Methods in the Social Sciences program, supporting instruction and student learning in quantitative methods.
Technology stack
ML & research: Python, PyTorch, TensorFlow, scikit-learn, Hugging Face, Jupyter, MLflow.
LLM & agent systems: LangGraph, LangChain, RAG, GraphRAG, MCP, tool-calling workflows, evaluation frameworks.
Graphs & data: Oracle Property Graph, RDF/OWL, SPARQL, PGQL, Neo4j, D3.js, SQL, Databricks.
Deployment: Docker, FastAPI, REST APIs, cloud AI platforms, CI/CD, monitoring and MLOps.