From research
to production.

Built a production multi-agent LangGraph system deployed on Google Cloud (GKE, Kubeflow) that turns an ambiguous, manual reconciliation workflow into structured, reviewable output. The pipeline ingests Excel and database records, resolves entities via fuzzy-matching, and classifies each one (MATCH / INSERT / ORPHAN / DISAGREEMENT) using an LLM (Kensho Spark / GPT-4.1) — with a human-in-the-loop checkpoint before anything is finalized.

Why it matters: Every output is auditable and nothing is finalized without sign-off — built and iterated directly with the business team using it daily in production.

Built with teammate Aaryaman Bajaj as an Applied ML final project: a semantic photo search system using CLIP and BLIP for natural language image queries. Enables users to search their photo library with descriptions like "sunset at the beach" or "my dog playing." Implemented zero-shot classification and cross-modal retrieval with FAISS for efficient similarity search.

Why it's impressive: Bridges vision and language modalities using state-of-the-art transformers. Handles semantic understanding, not just keyword matching. Production-ready with sub-200ms query latency on large image collections.

Built a DistilBERT binary classifier that routes each query to a lightweight or large model (Llama-3.1-8B) by predicted difficulty, cutting inference cost 61.6% while retaining 78.7% of large-model accuracy.

Validation: Tested the cost–quality tradeoff across 800+ benchmark samples using an evaluation framework built on hypothesis testing, cross-validation, and feature analysis (statistically significant, p < 0.01).

Built an autonomous agent that extracts, classifies, and stores information from unstructured text using Google Gemini API. Automatically processes documents, identifies key entities, and organizes data in Google Cloud Storage with intelligent tagging and retrieval mechanisms.

System design: End-to-end pipeline with text preprocessing, LLM-based classification, structured data extraction, and cloud storage integration. Deployed on Render with automated workflows.

AI-driven agricultural platform that recommends optimal crops, fertilizer strategies, and detects plant diseases using ML and deep learning. Built with Flask, PyTorch, and computer vision models. Processes soil data, weather patterns, and crop images to provide actionable insights.

Real-world impact: Helps farmers make data-driven decisions about crop selection and disease management. Combines traditional ML (for tabular data) with CNNs (for image classification).

AI-powered medical assistant using Retrieval-Augmented Generation (RAG) and Chainlit. Retrieves relevant medical information from knowledge base and generates contextual responses. Built with vector databases for semantic search and LLMs for natural language understanding.

Technical approach: Combines embedding models for semantic retrieval with LLMs for response generation. Ensures factual accuracy by grounding responses in retrieved medical documents.

Published research on extractive text summarization techniques. Investigated graph-based methods, sentence ranking algorithms, and semantic similarity measures for automatic summary generation. Compared approaches on benchmark datasets and proposed optimizations.

Research contribution: Demonstrated ability to conduct rigorous research, implement experiments, and communicate findings. Shows depth in NLP fundamentals and academic writing.

Built a backend data pipeline processing 200+ Treasury/FRED time series into a PostgreSQL warehouse, tuning the schema with caching and composite indexes for 40% faster historical queries. Layered an NLP sentiment pipeline over 100+ Fed speeches and used server-side aggregation to render 200+ series in a responsive analytics dashboard.

Architecture: Server-side aggregation and composite indexing were the key wins for keeping the dashboard responsive at scale — direct experience with financial market data in production.

Shipped a live GPS-based multiplayer game to the App Store with a 3-person team at Columbia Startup Studio — end-to-end ownership from build to live users, grown to 70+ downloads and 43 active users.

Team: Built with David Ge and Aaryaman Bajaj at Columbia Startup Studio — real-time GPS multiplayer mechanics shipped to production, not a class prototype.