I enjoy building machine learning systems and making sense of messy data, questioning assumptions, finding meaningful patterns, and shipping things people can actually use.
Building RepoChat, a Chrome extension that brings real-time collaboration directly inside GitHub. Users connect via GitHub ID, chat in personal or repo-linked groups, and get auto-posted issues and PRs streamed into conversations. Assign or claim tasks, reply in threaded discussions, create custom topic threads, and manage all pending and completed work from a dedicated MyTask view, without ever leaving GitHub.
Under active development · not yet publicly launched
Swipe to see more • Tap to expand
Contribution activity over the past year · @alexcj10
Fine-tuned Microsoft DeBERTa-v3 with LoRA on ~48K annotated sentences from the GMB corpus to perform token classification across 17 labels (8 entity types), achieving a 0.96 weighted F1 score. Built and deployed a production full-stack system featuring an async FastAPI backend on Hugging Face Spaces (Docker) and a responsive, glassmorphic React frontend on Vercel.
Analyzed 1M+ sales transactions across 75 global stores over 4 years to evaluate product performance, geographic distribution, and warranty reliability. Created a dimensional model in DuckDB, resolved formatting inconsistencies, and queried data to uncover insights like an 872% variance in volume between top and bottom performing countries. Highlighted tight margins in warranty claims across categories, confirming broad product reliability and identifying stable non-seasonal sales floors.
An end-to-end computer vision suite for fruit species identification, ripeness classification, and continuous shelf-life regression (estimating remaining days to expiry). Built using transfer learning with an EfficientNet backbone to model biological decay from surface features. Features a live web application, async Hugging Face inference API, and a distributable Python CLI tool.
Deep-dived into 99,441 orders from Olist, a Brazilian marketplace, to uncover revenue drivers, customer behavior, and operational performance. Analyzed over R$ 13M in total revenue across 27 states, unearthing São Paulo's dominance at R$ 5.2M, severe logistic bottlenecks in the North with delivery times nearing 30 days, and a widespread dependency on credit card installments averaging 2.85 per transaction. Identified top-performing product categories, evaluated seller concentration risks, and mapped order status distributions revealing a 97% delivery success rate.
K.E.S. Shroff College of Arts and Commerce, University of Mumbai. CGPA: 7.73/10
Python for Data Science | IBM (Cognitive Class), Dec 2023
Machine Learning Algorithms | Great Learning Academy, Nov 2023
Feel free to say hi at alexchandarjoshva@gmail.com.