Most of my work is in agentic RAG and deep learning, with some data engineering on the side. I've published a paper on hybrid RAG, I ship things that actually work, and I care more about what runs than what demos well.
Lets non-technical teams get answers from their data in plain English or Hinglish, without waiting on an analyst. Every query is checked for safety before it runs, so it can't break or change the database. Currently benchmarking on BIRD.
View GitHub →A RAG assistant that decides when to use its own memory and when to search the web, so answers stay current without wasting API calls. Reached 92% on a 25-question benchmark, and the basis of my published paper.
View GitHub →Helps platforms and newsrooms spot deepfake images before they spread. 94.05% accuracy, ROC-AUC 0.98, and fast enough for real-time use at under 12ms per image.
View GitHub →Gives disaster-response teams one live view of where earthquakes and bad weather overlap, so they know which rescue zones need extra planning. Updates every hour automatically.
View GitHub →Cuts hours of manual research down to one conversation. It searches, summarizes, and removes personal info, and saves progress so a crash doesn't lose the work.
View GitHub →Tells a business which customers are likely to leave, so the retention team can act before they do. ROC-AUC 0.836, with drift detection and monthly retraining so it stays accurate.
View GitHub →Lets anyone edit and find photos by just typing what they want, no Photoshop skills needed. Built on the Gemini API.
View GitHub →Automatically tags clothing images into 10 categories, the kind of thing e-commerce stores need for search and filters. Improved accuracy from 88.2% to 93.15%.
View GitHub →Proposes an explicit agentic retrieval decision layer that routes queries between vector memory, live web search, and a hybrid path, rather than retrieving blindly on every turn. Pairs a local Llama 3 with a Gemini 2.0 Flash fallback that takes over when the local model isn't available, and it saves useful facts from conversations on its own.
I'm a 21-year-old AI engineer from Himachal Pradesh, now based in Delhi. I did my BCA at Himachal Pradesh University and I'm finishing an Associate of Science in AI at Saras AI Institute, USA.
I learn best by building things end to end and watching where they break. That's how I've picked up most of what I know, from ML systems to research on agentic RAG.
Right now I'm looking for AI/ML internships or research-leaning junior roles where I can ship real systems and learn from people better than me.
Associate of Science in Artificial Intelligence
Saras AI Institute, USA · 2025 – 2026
Machine Learning, Applied Generative AI, Math for AI, Cloud AI deployment
Bachelor of Computer Applications (BCA)
MCM DAV College · Himachal Pradesh University · Completed 2026
Built my CS foundations in data structures, DBMS, operating systems, networks, and statistics.
Moved into AI during my BCA to focus on building systems that solve real business problems. Coursework in Machine Learning, Applied Generative AI, Math for AI, Data Engineering, and cloud AI deployment.
Built 8+ AI projects across RAG, computer vision, agents, and data engineering, each aimed at a real use case like spotting deepfake images, predicting customer churn, and supporting disaster response. Highlights include a RAG assistant at 92% accuracy and a deepfake detector that works in real time.
Published Moksh AI v3.0, an agentic retrieval decision layer that lifted benchmark accuracy from 64% to 92% over a naive-retrieval baseline.
Building a tool that lets non-technical teams ask questions about their data in plain English or Hinglish. The LLM only writes SQL, and a validator checks every query before it runs. Evaluating it on the BIRD benchmark to measure how much validation cuts wrong answers.
Merging live earthquake data with weather conditions to flag when rescue efforts could be complicated, plus the three things that actually broke along the way.
Read article →
I've been a Python person my whole ML journey, until I looked at where Rust already sits in the stack I use daily. Tokenizers, dataframes, agent runtimes: it's already there.
Read article →
The first time I got an LLM to turn plain English into working SQL, I felt like a wizard for about ten minutes. Then the boring questions started: what if it writes a DELETE, or returns a clean number that is simply wrong?
Read article →If you're looking for a passionate AI engineer, let's connect.