What I've built

Projects.

2025 Featured

Hierarchical Video RAG Pipeline

Problem

Long-form YouTube content is nearly impossible to navigate or study from — raw transcripts blow past context windows and key insights stay buried.

Approach

Built a batched hierarchical summarization pipeline that compresses transcripts in stages, then layered retrieval-augmented Q&A on top with a Streamlit interface for notes, flashcards, and multi-video queries.

Impact
  • Queryable knowledge base from multi-hour playlists
  • Deployed live — in active use as a study tool
  • Context-window constraints solved via staged compression
LLM PipelineRAGStreamlit
Source Live App
2025

ResNet & Attention LSTM Image Captioner

Problem

Extracting structured, human-readable descriptions from image-heavy inputs requires joint vision and language modeling — not just classification.

Approach

Trained an end-to-end captioning pipeline: ResNet-50 encoder with transfer learning feeding an attention-based LSTM decoder, implemented from scratch through IIT Bombay's Seasons of Code program.

Impact
  • Working vision-to-text pipeline trained end-to-end
  • Attention weights give interpretable image-text alignment
  • Full training + inference code released on GitHub
PyTorchResNet-50Attention LSTM
Source
2026

LangGraph Stateful Multi-Agent System

Problem

Single-prompt LLM calls collapse on tasks that need decomposition, delegation, and memory across multiple reasoning steps.

Approach

Designed a stateful graph orchestration with LangGraph and Google ADK: modular agents for task decomposition, routing, and response generation, with explicit state passing between nodes.

Impact
  • Reliable multi-step task delegation across specialized agents
  • Stateful conversations with controllable control flow
  • Documented failure modes of agent reliability & hallucination handling
LangGraphMulti-AgentGoogle ADK
Source
2025

Large-Scale Market Basket Association Analytics

Problem

Inventory decisions at scale were guesswork — no systematic view of which product combinations actually co-occur across millions of transactions.

Approach

Engineered an association-rule mining pipeline over 3.4M+ orders with pandas, tuning support/confidence thresholds to surface statistically robust bundles instead of noise.

Impact
  • 3.4M+ orders mined for repeatable product combinations
  • 15+ dashboard outputs guiding inventory planning
  • Rule quality controlled via thresholded support/confidence
PandasAssociation RulesBI
Source
2025

Campus-Scale Sports & Operations Portal

Problem

13,000+ IIT Bombay students had no centralized hub for tournament schedules, fixtures, and sports contacts — updates lived in scattered group chats.

Approach

Designed and built the official institute sports portal (React frontend, Django backend) as Institute Web Convenor, focusing on responsive UX and load performance.

Impact
  • 13k+ students served campus-wide
  • ~30% faster page loads vs. prior setup
  • Single source of truth for events, schedules, contacts
ReactDjangoResponsive
Source