Minimal online presence.

Repo: performantvideo

Faster scrubbing controls when viewing videos.

• frame.io and other web video players struggle with buffering when playing forwards at high speeds or backwards. Building a method for better performance.
• Encoding a reversed version of the video with a circular buffer of rendered frames for instant backward navigation without re-decoding
• Mipmap-style pre-computed versions of the video at faster playback rates for lag-free display

Repo: orbit

A pipeline execution framework built for a monolith infrastructure.

• Demo video showing it off!
• Operates in your TypeScript backend, using type-safe handlers instead of YAML configs. A library to import rather than a platform to deploy.
• Built-in and configurable behavior for automatic retry logic, state persistence, crash recovery, background workers, and real-time performance monitoring
• CLI functionality
• Web dashboard interface for monitoring and management

Repo: CanvasPlanner

Learning about products like Figma by building an online multi-user planning tool.

• Demo video explaining how it works!
• Built-in graph-based time planner to estimate workflow duration
• Exploring sync engines, websockets, and multiplayer
• Uses a node-based interface, built with react-flow, convex, and Socket.IO
• Client-side caching to prevent unnecessary rerenders

(Private repos for security purposes)

Managing data collection and reporting for service organizations.

• Architected and built full-stack CRM platform from zero-to-one, owning end-to-end technical decisions
• Designed multi-tenant backend coordinating data across third-party APIs and internal microservices
• Handled all design, implementation, migration of production databases (PostgreSQL, MongoDB)
• Built and moved across multiple providers (AWS, GCP, DigitalOcean)
• Implemented SOC 2/HIPAA compliance, secure auth and data storage
• Building frontend
• Pre-pivot, worked on software facilitating ESL program management and in-class instruction

Repo: Cathedral

Applying RL methods to a simple but deep MDP board game with an absurd state space.

• Experimenting with MCTS, Deep-Q Learning, and Approximate Q-Learning

Multiplayer board game developed with Nick and Ben Howe

• **Box and pieces
• Merchant game set in space, featuring player interaction without hostility
• Playtested and collected feedback, rebalancing and reworking mechanics
• Edited rulebook

(Private repos for security purposes)

LinkedIn projects write-up available!

• Religion Library DB, Lycium DB, Hollaway Classics Library DB, Binary Worlds Lab Website
• Currently in production on Amherst servers
• Developed full-stack, and built REST APIs (TypeScript, React, Next.js, Node.js, Flask, PostgreSQL)
• Designed scalable, reliable database architecture handling concurrent access and complex query workloads
• Working with department members for requirements gathering and analysis

Completed my math major! Focus on Galois and Group Theory.

Repo: laysumm | HuggingFace: PEGembed

For the grad course Advanced NLP w/Prof. Mohit Iyyer, developed over 4 months with the help of Zhiheng Wang, Jiarui Liu, and Ahmed Jaafar.

• Researched and developed fine-tuned generative AI model
• Adapted Seungwon Kim's method (2020) of extractive summary, followed by abstractive summary
• Extractive summary using cosine similarity to extract top N sentences
• Pre-trained PEGASUS model, fine-tuned on the eLife dataset
• Met baseline models on metrics BERTScore and METEOR, while outperforming all baselines on ROUGE-2 (33%), ROUGE-L (35%), and Human Lay Readability (21%)

Repo: Anisotropic-Total-Variation

Intel-funded RPA startup leveraging computer vision models for automated app testing.

• Creating AWS pipelines for handling training and model data during the fine-tuning of these models
• Building monitoring and error recovery for pipeline robustness
• Training models using XGBoost
• Modularizing and refactoring various code

Team led by Prof. Lee Spector, centered around the Push family of genetic programming languages.

• Created a visualizer in Clojure to visually present model execution, in hopes of understanding AI black-box behavior

I like ML, fullstack development, and databases.