Data Science @ UCSD

Building multi-agent systems that think, decide, and execute.
Autonomous agents that actually work in production.

7 Quarters Provost Honors
3 Competition Wins
90+ Students Taught
Ishaan Gosain
Scroll to explore

About

Hi, I'm Ishaan—I build ML systems, generative AI apps, and agentic workflows. Most of my time goes into making things that actually work, not just things that demo well.

My interests span probabilistic modeling, NLP, recommendation engines, and increasingly, anything involving LLMs and autonomous agents. I also tutor and mentor—turns out teaching is the fastest way to realize what you don't actually understand.

When I'm not coding, you'll find me at the gym, losing hours to FIFA or story-driven games, or asking seniors and professors too many questions about this field.

Currently looking for opportunities where I can build real systems, not just proof of concepts.

Currently

  • Data Science @ UCSD (Graduating 2027)
  • ML Tutor for DSC 140A (90+ students)
  • FIFA Weekend League competitor
  • Story-driven game enthusiast

Previously

  • Data Science Intern @ Accolite
  • Projects Mentor @ ACM AI (ended Sept 2025)
  • Data Engineering @ GSPANN
  • State-level soccer, Delhi
  • IIT JEE qualified (top 2% of 1.5M+)

Recognition

  • DS3 Winter Cohort — 1st Place (Music Recommendation)
  • DS3 Summer Cohort — 3rd Place (ML Under the Hood)
  • DSC 106 — Best Project & People's Choice
  • Provost Honors — 6 consecutive quarters
  • 3.96 GPA

Selected Work

Projects that taught me something new

2025 Personal Project

Agentic AI Platform – Multi-Agent Database Query System

The Problem: Querying databases requires SQL knowledge, limiting accessibility for non-technical users. Traditional solutions either lack flexibility or require extensive training.
The Solution: Built a production-ready multi-agent system that transforms natural language questions into structured database queries. The platform uses LangChain agents orchestrated by LangGraph workflows, with MCP (Model Context Protocol) for seamless tool integration.
The Impact: Intelligent query routing, result caching with Redis (reducing query time by 60%), and comprehensive monitoring with Prometheus metrics. Deployed on Railway.app with full observability.
Python LangChain LangGraph MCP FastAPI Redis Prometheus Railway.app Streamlit
Live Demo Code
Agentic AI Platform Interface
2025 3rd Place — DS3 Summer Cohort

ML Under the Hood

The Problem: ML algorithms are often taught as black boxes. Students struggle to understand how K-means converges, why PCA reduces dimensions, or how EM algorithms work step-by-step.
The Solution: Built an interactive educational platform with real-time algorithm execution visualizations. React + D3.js frontend shows step-by-step tracing, parameter adjustment, and 2D/3D visualizations for K-Means, PCA, Linear Regression, and EM algorithms.
The Impact: Won 3rd place for Best Project & Presentation at DS3 Summer Cohort. Deployed with FastAPI backend on Render and actively used in UCSD tutoring sessions, helping 90+ students visualize ML concepts in real-time.
React TypeScript D3.js Plotly.js FastAPI Python Render
Live Demo GitHub
ML Under the Hood K-Means Visualization
2025 1st Place — DS3 Winter Cohort

Mood-Based Music Recommendation

The Problem: Traditional music recommendation systems ignore emotional context. Users want music that matches their current mood, not just their listening history.
The Solution: Built an emotion-aware recommendation system analyzing YouTube comments for sentiment. Multi-stage pipeline: RoBERTa transformer for 11-dimensional emotion extraction, K-means clustering for mood categorization, and XGBoost ensemble for personalized recommendations.
The Impact: Achieved 10x performance improvement (200ms→20ms response time) and won 1st place at DS3 Winter Cohort. Deployed on AWS Elastic Beanstalk with full CI/CD, processing 2,000+ songs with 89% user satisfaction.
Python RoBERTa XGBoost Flask Pandas Scikit-learn AWS Elastic Beanstalk
Live Demo Code
Music Mood Recommender Interface

Other Work

Study Buddy AI

Cloud-native AI study assistant with Kubernetes orchestration and CI/CD pipelines. Tech: Python, LangChain, Docker, Kubernetes, Jenkins, ArgoCD

GitHub

Earthquake Impact Visualization

Interactive 3D visualization showing infrastructure's role in disaster impact. Won Best Project & People's Choice in DSC 106 (top 8% of 182 students). Tech: D3.js, JavaScript, HTML5, CSS3

Live Demo

AI Customer Support Automation

End-to-end GenAI customer support system that automated 60% of routine inquiries, reducing response times by 40%. Tech: Python, GPT-4, n8n, Supabase Vector DB, Docker, RAG

GitHub

Professional Journey

From data pipelines to teaching ML—each role taught me something new

Sep 2025 – Dec 2025 Completed

Instructional Assistant / Tutor

Halicioglu Data Science Institute, UC San Diego San Diego, CA

Teaching upper division Probabilistic Modeling and Machine Learning (DSC 140A) to 90+ students under Professor Tevfik Berk Ustun. Making complex concepts like gradient descent, regularization, and EM algorithms actually make sense.

  • Conduct weekly office hours and exam review sessions
  • Guide students through regression, classification, and generative models
  • Help students debug ML implementations and understand theory
Key Learnings: Teaching forces you to understand fundamentals deeply. Explaining gradient descent to beginners revealed gaps in my own understanding. The best way to master a concept is to teach it.
Teaching Machine Learning Python Scikit-learn
Jun 2025 – Sep 2025 Completed

Gen AI Automation Intern

Accolite Digital (acquired by Bounteous) Gurgaon, India

Global digital transformation consultancy with 5,000+ employees serving Fortune 500 clients. Built production-ready GenAI systems for enterprise customer support automation.

  • Built end-to-end GenAI customer support automation using n8n, GPT-4, and Supabase Vector DB (RAG), handling 60% of routine inquiries
  • Reduced average handle time by 40% and improved first-contact resolution by 15% through semantic search and sentiment-aware personalization
  • Deployed modular, GDPR-compliant architecture using containerized microservices with Docker
Key Learnings: Production ML systems need monitoring from day one. Enterprise deployments require GDPR compliance and modular architecture. Semantic search and sentiment analysis can dramatically improve user experience when combined thoughtfully.
GPT-4 n8n Supabase Vector DB Docker RAG
Summer 2025 Completed

Projects Mentor

ACM AI @ UCSD San Diego, CA

Mentored 4 students completely new to AI/ML through a complete machine learning project lifecycle. Led weekly meetings and mentoring sessions, teaching end-to-end ML development from data acquisition to deployment. The team built an Airfoil Self-Noise Prediction system achieving R² = 0.949 using XGBoost with advanced techniques like SHAP analysis and Bayesian optimization.

  • Taught complete ML lifecycle: data acquisition, data engineering, EDA, baseline models, model selection, and hyperparameter tuning
  • Led weekly technical sessions covering advanced techniques including SHAP analysis, Bayesian optimization, and cross-validation
  • Guided team to achieve R² = 0.949 with XGBoost, resulting in a production-ready project with comprehensive documentation
Key Learnings: Mentoring beginners taught me to explain complex concepts simply. Breaking down the ML lifecycle into clear steps helped me understand the process better. Leading a team requires balancing technical depth with accessibility.
Mentoring XGBoost SHAP Analysis Bayesian Optimization Team Leadership ML Lifecycle
Jun 2024 – Sep 2024 Completed

Data Science Intern

GSPANN Technologies Remote

Digital solutions company specializing in e-commerce and retail technology for enterprise clients. Built RAG-powered systems for resume screening with real-time analytics.

  • Built RAG-powered resume screening system using LangChain and FAISS, processing 10,000+ resumes with 2-5 second query response
  • Implemented RAG Fusion for multi-perspective candidate matching, improving retrieval accuracy over standard semantic search
  • Deployed Streamlit interface with real-time analytics dashboard and JSON export capabilities using OpenAI GPT-4
Key Learnings: RAG Fusion significantly outperforms standard semantic search for multi-perspective queries. Vector databases like FAISS are crucial for production-scale retrieval. Building real-time analytics requires careful consideration of query performance and user experience.
LangChain FAISS RAG Fusion Streamlit GPT-4
Live Demo ↗

Tech Stack

Languages

Python, Java, JavaScript, SQL

ML/AI

PyTorch, TensorFlow, Scikit-learn, XGBoost, Hugging Face, LangChain

Data

Pandas, NumPy, Spark, D3.js, Streamlit

Cloud/DevOps

AWS, Docker, Kubernetes, Git, n8n

Let's Talk

Want to chat about ML, data science, or building systems that actually work?

Email ishaangosain08@gmail.com LinkedIn Connect professionally GitHub Check out the code