Support Agent
A router sends each question to an orders agent or a read-only policy agent. Large refunds wait for a human.
Routing accuracy 90% → 95% on 20 test questions
- Python
- FastAPI
- Pinecone
- Docker
Things I've built
Resume Chatbot
Ask questions about my experience, and get answers from my own documents.
In progress
- RAG
- LangChain
- streaming
Text-to-SQL Evaluation
A reproducibility study showing exact-match scoring badly underestimates LLM performance.
0.0 exact match vs ~83% valid SQL on BIRD
- LLMs
- evaluation
- OpenAI API
Lung Segmentation
U-Net and Attention U-Net on chest X-rays, tested under noise, blur, and contrast loss.
U-Net: Dice 0.87 on clean images. Noise-trained Attention U-Net: 0.84 under low contrast
- Deep learning
- segmentation
Emoji-Aware Sentiment
Do emojis improve sentiment classification? Three strategies across classical and neural models.
Emojis changed accuracy by 0.001
- NLP
- scikit-learn
- BiLSTM
Stroke Prediction
Class-imbalance handling for a rare-event medical dataset.
Upsampling lifted recall to ~0.80 to 0.85, but precision stayed low (~0.12 to 0.14)
- R
- imbalance
- evaluation
ASK MY RESUME
Chat with my resume
An AI assistant that answers only from my resume and projects. It can make mistakes, so check anything important with me.