Deeptanshu Lal

A System One model for multilingual toxicity: it reads a comment once and returns six typed probabilities, called as a function inside software rather than prompted like a chat model. Out-ranks every public multilingual baseline (toxic AUC 0.992 vs Detoxify's 0.969) at 11.8 ms a comment.

Recommendation engine built around a hybrid-weighted similarity that fuses TF-IDF ingredient vectors with structured recipe metadata. It needs no ratings, so it works from the very first query: the cold-start case where collaborative filtering has nothing to go on. Indexes 7,000+ recipes.

Real-time contribution tracker built for 200+ developers during Hacktoberfest 2024. A GitHub webhook stream processes 1,000+ PR events with rate limiting and auto-ranking updates. Deployed on AWS EC2 behind Nginx, and held 99.99% uptime under concurrent load throughout the event.

Ranks a batch of resumes against a job description. Pulls skills, experience, education and contact details out of each PDF or text resume, then scores skill fit with BERT embeddings so a close skill still earns partial credit.