Projects

Tech stack: LangGraph, PyTorch, Python, MongoDB, Gemini LLM, Voyage AI, Cohere Reranker, DVC, Prefect, MLFlow, Pinecone

Engineered a two-tower recommendation system using Feast-backed features and holdout-based offline evaluation, outperforming a cosine-similarity baseline by 6.6× on Recall@10, 3.8× on NDCG@10, and 4× on Hit@10, while reducing inference latency from 63.7ms to 2.2ms.
Developed a LangGraph multi-agent system for plant Q&A, side-by-side comparison, exploring recommendations, and live web search via Tavily, supported by a hybrid RAG pipeline with cosine similarity, BM25 reranking, and recursive/sliding window chunking.
Designed a synthetic data generation framework that simulated user–plant interactions for training and evaluation in a data-scarce domain, successfully capturing 8 out of 10 target behavioral patterns.
Containerized the FastAPI backend with Docker and deployed on Railway and Vercel, serving a catalogue of 1000+ plants.

Team Lead

Tech stack: LAMP stack, Flask, AWS EC2, Google Gemini

Led a team of 6 to develop a search platform for 100+ AI tools using keyword or NLP-based search, winning 1st place out of 12 teams in a software engineering competition.
Designed and optimized a relational database schema and search pipeline using SQL, enabling natural language queries with ~0.4s average latency.
Collaborated cross-functionally with designers, frontend and backend engineers, and QA in an Agile environment to improve team productivity and workflow efficiency.

Tech stack: Keras, Python, OpenCV, ResNet50, Scikit-learn, Matplotlib, SciPy, scikit-image, TensorFlow

Designed a computer vision pipeline with OpenCV to detect, warp, and segment real-life chessboard images into 64 cropped squares for piece classification.
Fine-tuned a ResNet50 model on a custom dataset of 2,000+ images, achieving 94% classification accuracy across all classes.
Developed an adaptive thresholding system using linear regression to handle variable lighting and image conditions, improving successful board detection rate by 4× across 150 diverse real-world unseen board images.