Projects

01
LangGraph System Design React Neo4j / Cypher Claude API FastAPI Data Modeling

Deloitte-Mentored Capstone

Personalized Supplement Recommender

81% accuracy
329K graph entities
3.4M relationships
33% lower latency

A multi-agent system that flags dangerous supplement–medication interactions and generates safe, personalized recommendations grounded in a biomedical knowledge graph.

Built a knowledge graph-driven AI system on Neo4j where an orchestrator coordinates specialist agents for entity extraction, normalization, retrieval, and synthesis - enabling evidence-based reasoning at scale.
Restructured a linear pipeline into conditional LangGraph orchestration, cutting response latency 33% through targeted agent routing; deployed via React + Vercel.
Grounded LLM-generated Cypher in the live database schema with dry-run validation, blocking hallucinated fields before execution.
Implemented JSON logging and guardrails across a stateful 7-agent workflow - zero unsafe outputs across 62 golden-set evaluations.
View on GitHub ↗ View Website ↗ View Live Demo ↗
02
RAG Hybrid Search Reranking LLM Evaluation HydraDB FastAPI React

RAG · Evaluation Testbed

Stock Market Decoder

90+ evaluation cases
9 test suites
13 SEC filings

A dashboard that explains why a stock moved - mapping monthly price history to the SEC filings and press releases behind it, with every answer traced back to a cited source span.

Built an LLM-routed retrieval pipeline pairing hybrid search and reranking with token-efficient context assembly, so every answer is grounded in a citable source span rather than model memory.
Engineered an evaluation harness of 90+ cases across 9 suites, grading routing, sources, and answers at both trajectory and task level.
Designed an ingestion and indexing schema that splits embedded text from filterable metadata, enabling scoped graph-augmented retrieval.
View on GitHub ↗
03
Prompt Engineering Black-Box Evaluation Benchmarking (AdvBench) Qwen LLMs

AI SAFETY · ADVERSARIAL ML

Jailbreaking Large Language Model

78% success rate
6.2 queries-to-success
25 adversarial objectives

A system that tests how and when language models break under adversarial prompting.

Built an iterative PAIR (Prompt Automatic Iterative Refinement) pipeline to probe LLM safety through attacker–target model interactions
Evaluated Llama-based models under adversarial pressure, measuring success rates and query efficiency across structured attack scenarios
Designed a feedback-driven refinement loop to adapt prompts based on prior responses and improve attack effectiveness.
Closed Source
04
Transformer Architecture Tokenization Transfer Learning Streamlit Text Classification

NLP · Classification

MBTI Personality Classifier

94.8% train accuracy
87% validation accuracy
+12% F1 improvement

A natural language processing system that predicts personality type from text.

Fine-tuned a BERT-based NLP classifier with Hugging Face and TensorFlow - improved F1-score by 12% through optimized tokenization and hyperparameter tuning.
Built an interactive Streamlit app serving real-time personality predictions across 1K+ user inputs with strong generalization on unseen data.
View on GitHub ↗
05
Cross-Validation Model Evaluation Feature Engineering Scikit-Learn

Machine Learning

Credit Card Fraud Detection

0.99 ROC-AUC
0.93 PR-AUC
−18% false negatives

A machine learning system that identifies fraudulent transactions in financial data.

Processed 284K anonymized transactions with a 0.172% fraud rate - addressed severe class imbalance using SMOTE and stratified sampling.
Trained and cross-validated Logistic Regression, Random Forest, XGBoost, and Isolation Forest models with rigorous evaluation.
Optimized classification thresholds via precision-recall analysis, reaching 0.93 PR-AUC and cutting false negatives by 18% - where misses carry real cost.
View on GitHub ↗