Projects

Subject-wise bunk planning and low-attendance recovery using greedy-algorithm logic.

Dual-mode engine — Bunk Planner (safe skips) and Recovery Tracker (classes needed to recover attendance) — with live recalculation on every input.

Deployed as a PWA on Vercel: installable on mobile, localStorage persistence, zero login friction.

Private PostHog dashboard tracks DAU/MAU and college-wise usage.

End-to-end ML pipeline classifying financial text sentiment (positive/negative/neutral) on a combined FiQA + Financial PhraseBank corpus (6,051 entries post-deduplication).

Custom 900+ term financial sentiment lexicon built to resolve domain-specific ambiguity (e.g. "debt," "liability") that generic NLP misreads, concatenated with TF-IDF features.

Benchmarked 5 algorithms (Logistic Regression, Random Forest, Naive Bayes, KNN, Neural Network) with SMOTE and class-weighting for imbalance — Logistic Regression performed best at 0.64 macro-F1.

Expanded into a full-stack deployed app with candlestick charting and a live stock explorer.

AI-driven training environment automating IT support ticket handling end-to-end, no hardcoded logic.

An agent classifies incoming tickets, assigns priority, and routes them to the correct team, improving through reinforcement learning (reward on correct routing, penalty on incorrect) — simulating a real-world IT workflow.

Built and submitted a complete AI social concierge agent for Google's official Intensive AI Workshop.

Applied multi-step reasoning, context-aware memory, and structured prompt engineering using Gemini models.

Designed and documented the full agent architecture independently with Streamlit UI and SQLite memory store.