AI Engineer building production-deployed Generative AI and Machine Learning systems.
Currently looking for entry-level AI/ML/GenAI Engineer roles in Toronto/GTA Ask me about LangChain, RAG, multi-agent systems, FastAPI, or ML pipelines Reach me: niharikabisoyi28@gmail.com
| Project | Live Demo |
|---|---|
| AI Customer Review Intelligence System | Open Live API → |
| AI Travel Planner (Multi-Agent) | Open Live App → |
- Generative AI systems — RAG pipelines, multi-agent workflows (LangChain & LangGraph), LLM-powered apps
- Machine Learning pipelines — end-to-end, from data cleaning to deployment
- Production backends — FastAPI + MongoDB, with logging, error handling, and automated tests
- Cloud deployment — Render and Google Cloud Platform (Cloud Run, Cloud Storage)
Languages: Python, SQL
Generative AI: LangChain, LangGraph, RAG, OpenAI API, ChromaDB, embeddings, prompt engineering, multi-agent systems
Machine Learning: Scikit-learn, LightGBM, CatBoost, SMOTE, feature engineering, hyperparameter tuning
Backend & Data: FastAPI, MongoDB, Pandas, NumPy, Pydantic, Loguru, pytest
Cloud & DevOps: Google Cloud Platform (Cloud Run, Cloud Storage), Render, Git, GitHub
Visualization: Streamlit, Matplotlib, Seaborn
Live API that analyzes customer reviews in real time using AI
Send any review → get back sentiment, a summary, key issues, and recommendations — instantly. Built with FastAPI, OpenAI, MongoDB, and ChromaDB for semantic search (finds reviews by meaning, not just keywords). Deployed live with automated tests and structured logging.
An AI travel assistant powered by 4 specialized agents
Ask it to plan a trip, find places, or recommend activities — a multi-agent system built with LangGraph routes each request to the right specialist agent. Combines real-time Google Places data with an AI knowledge base (RAG). Deployed on Google Cloud Run.
An AI that runs mock interviews and gives feedback
Generates relevant interview questions from a knowledge base, asks adaptive follow-ups based on your answers, and scores responses for accuracy and depth. Built with LangChain, OpenAI/Mistral, and ChromaDB.
Predicting which customers are about to leave
Improved recall from 49% to 75% by identifying and fixing a class imbalance problem (80:20 split) using SMOTE and tuning an SVM model — meaning the model now catches far more at-risk customers, directly reducing lost revenue. Benchmarked against 4 models. Deployed with Streamlit.
Estimating resale prices from car details
Achieved 83% R² using LightGBM after cleaning messy data — fixing extreme price outliers, missing values, and a highly skewed target variable. Trained model now loads from Google Cloud Storage at runtime. Deployed with Streamlit.
- Adding streaming responses and automated LLM evaluation (RAGAS) to my GenAI projects
- Expanding cloud skills with BigQuery and Vertex AI
- Applying CI/CD and model versioning practices across all projects