Authored and published a Claude Code skill that intercepts 11 distinct 'cowardly' patterns in Claude's behaviour — premature skipping, three-alternative hedging, workaround-as-solution, scope shrinking, error-avoidance theatre, and others — and rewrites Claude's response strategy to commit to root-cause fixes. Designed the skill so Claude self-calls out slippages in real time (e.g. 'I was about to suggest skipping this — let me actually solve it'). Benchmarked against baseline Claude across 3 realistic scenarios (Selenium form-handling, Cloudflare-protected scraping, intermittent-crash debugging); the debugging eval moved from 2/5 to 5/5, with the largest deltas on the hardest tasks. Distributed via the npx skills CLI, manual git clone, and a portable .skill bundle.

Built a multilingual classification pipeline over 79,081 samples in 11 languages to detect AI- vs. human-written text, training classical models (Logistic Regression, Random Forest, SVM) on TF-IDF and bag-of-words features to reach 94.9% F1; fine-tuned XLM-RoBERTa-base to attribute text to 8 popular LLMs (e.g., GPT-4, Vicuna-13B) with 86.1% macro F1 and 86% accuracy; integrated SHAP and LIME to generate 100+ token-level explanations for cross-lingual transparency.

Collaborated with IT department's assistant head on a research project wherein devised a system utilizing YOLOv5 to detect pistols in real-time from CCTV, it sends an alert to concerned authority upon detection. Authored and presented a paper on same at ICACTA conference in October 2023 with securing 2nd best paper of category

Collaborated with C-DAC India to help XII standard CBSE board students understand derivation of Sum of N squared formula implementing 3D cubes and an interactive User Interface. Engineered web application by synthesizing Next.js, React3D, and Tailwind CSS technologies