Shreyan Paliwal

Undergraduate at Harvard University

GitHub·LinkedIn

About

I am a diehard Baltimore Ravens fan (yes, it's been a rough couple years. yes, I still think they're going to win the Super Bowl). When I'm not arguing about how many times Lamar Jackson has been robbed of an MVP, I'm studying math, computer science, and statistics at Harvard, exploring sports analytics and machine learning research.

Experience

Spring Silicon (Opt32)

Research & Engineering Intern

2026

Jane Street

First Year Trading & Technology Program (FTTP)

2026

Optiver

FutureFocus Program

2026
2025 – Present

Harvard Undergraduate Quantitative Traders

Trader

2025 – Present

Publications

Cache You Later: Post-Compression KV Repair for Long-Context Agentic LLM Inference

ICML 2026 Workshop on Adaptive Foundation Models (AdaptFM)OpenReview

First-author paper introducing RepairKV, a runtime operator that restores previously evicted key-value tokens to an LLM's active cache during a multi-turn agentic session. It argues that the KV cache should be multi-tiered and governed by a two-way mechanism, i.e. both eviction and repair rather than just eviction, so tokens dropped early can be reinstated once a later turn reveals they matter, rather than being lost for good. On a four-query needle-in-a-haystack benchmark, recovering specific facts buried in a long, mostly-irrelevant context, repairing just 96 tokens (0.3% of the full context) lifts retrieval from 24.5% to 91.0% at the same active-cache budget.

Projects

GridVeda

TreeHacks 2026 Winner

AI-powered grid intelligence platform for real-time transformer fault monitoring. Built an ensemble ML pipeline with a variational quantum circuit achieving 99% fault classification accuracy. Won one of 8 main sponsor tracks (Sustainability, Best Prototyping Process) at Stanford's premier hackathon (1,000 participants from 15,000+ applicants).

Chaos Game

Visualization of the Chaos Game, generating fractal patterns by iteratively plotting points using stochastic affine transformations on vertices of a polygon.

NFL Coverage Analytics

Developed novel receiver and cornerback evaluation metrics using proprietary tracking data. Built context-adjusted models that isolate individual skill in man-to-man coverage, controlling for route design, defensive scheme, and game situation. Research currently in use by and IP of an NFL team.

Awards

USACO Platinum Division Top 100

USAJMO Qualifier

Harvard ICPC 3rd Place Individual