About
Data science degree, engineering instincts.
I'm Gurleen Singh Bhatia, an AI engineer and MSc Data Science student at Chandigarh University, focused on agentic workflows, LangGraph orchestration, RAG architectures and Model Context Protocol integration.
My work sits at the seam between research ideas and production reality: LangGraph state machines routing real support tickets, local-first RAG pipelines with no data egress, and FastAPI services with JWT auth, Redis, PostgreSQL and Docker Compose so someone other than me can actually run them.
I'm also a published researcher on sub-10B parameter autonomous agents — the same restraint runs through everything I build: an agent that resolves most cases and escalates the rest with full context beats one that answers everything and is quietly wrong.
Gurleen Singh Bhatia
AI Engineer — Agentic Systems, LLMs & RAG
Publication
Unlocking the Potential of Small-to-Medium LLMs in Autonomous Agent Architectures: A Comparative Analysis of Gemini, Llama, and Qwen
Gurleen Singh, Rizwan Yousuf — GRENZE International Journal of Engineering and Technology (GIJET), Vol. 12 (2026), Issue 2, pp. 5666–5672
- — Evaluated sub-10B models (Llama 3.1 8B, Qwen 2.5 7B, Gemini 3.1 Flash Lite) inside the Model Context Protocol framework.
- — Co-developed “Complete-MCP”, an evaluation pipeline for single-step, sequential and multi-hop reasoning in simulated enterprise environments.
- — Showed >80% task completion for local sub-10B models with adaptive parsing and heuristic rate-limiting.
Education
MSc Data Science
Chandigarh University · Jul 2024 – Jun 2026 · CGPA 7.35
Data Structures & Algorithms, Statistics, Linear Algebra, Machine Learning — paired with a self-directed focus on LLM systems engineering.
BCA — Computer Applications
Arka Jain University · Aug 2021 – May 2024 · CGPA 8.70
Python Programming, Deep Learning, Neural Networks, Database Management Systems.
Skills
The toolkit
Agentic AI
- LangGraph
- LangChain
- Model Context Protocol (MCP)
- Multi-agent systems
- Semantic routing
- Human-in-the-loop (interrupt())
LLMs & retrieval
- RAG architectures
- Google Gemini 2.5 / 3.1
- Groq API
- Ollama / LM Studio (local)
- Google GenAI SDK
Backend & data
- Python (AsyncIO, Pydantic)
- FastAPI & REST
- PyJWT / OAuth sessions
- PostgreSQL, SQLite
- Redis Pub/Sub & caching
DS & DevOps
- Scikit-Learn, Pandas, NumPy
- NetworkX, Matplotlib
- Docker & Docker Compose
- Render, Streamlit Cloud
- uv, Git, Linux
Certifications & internships
Credentials
Virtual internships (Forage)
- BCG X — Data Science Job Simulation (EDA, feature engineering, predictive modeling)
- Accenture North America — Data Analytics & Visualization
- Accenture North America — Product Design Virtual Experience
- Moreton Bay Regional Council — Web Development Virtual Experience
Technical certifications & badges
- Google Cloud — Generative AI Fundamentals
- Google Cloud — Responsible AI: Applying AI Principles
- Google Cloud — Introduction to Large Language Models
- Google Cloud — Generative AI Explorer: Agent Platform
- Google Cloud — Introduction to Vertex AI Studio
- Kaggle — Machine Learning · Pandas Data Analysis
- IBM — Python for Data Science
How I work
Four steps, in this order, every time.
Understand the failure mode
Before any model choice, I work out where the current workflow actually breaks and what an unacceptable wrong answer looks like.
Design the graph, not the prompt
Routing, handoffs and fallbacks become explicit topology. If behaviour matters, it belongs in the architecture — not buried in a prompt.
Gate on confidence
Every autonomous path gets a threshold and an escape hatch. Systems that escalate well beat systems that guess confidently.
Ship something usable
A dashboard, an API and a Docker Compose file — so the work is reproducible and reviewable, not a notebook screenshot.