I graduated high school a year early in June. I’m spending this year doing research, building projects, and developing my skills while applying to college.
For Same Dynamics, Different Structures, I built experiments showing how different models can produce the same motion, then tested ways to distinguish them.
Accepted poster · NeurIPS AI for Science ’26 Accepted · IEEE MIT URTC ’26Explore Same Dynamics ↗
I coauthored ROOK, implementing the optimizer, running comparisons, and profiling its training computation.
I co-founded Code4Cause, now with 100+ chapters in 8+ countries. I designed and taught GreenSight’s four-week ML curriculum; students shared their research with local city council members. The organization also helped women in Afghanistan set up digital wallets to receive income.
I worked on nine cars and two motorcycles, sold ten, and gave one to a friend. Over three months, I earned about $10,000 after purchase costs and parts.
I organized a three-day hacker house for 19 people, selected from 180 applicants. I handled sponsor outreach, participant selection, food, lodging, and programming, securing about $15,000 in support from sponsors including Tether and ElevenLabs.
Two attendees began working toward a startup together; others started research collaborations.
I led FAI’s state team, helped raise $50,000+, ran a stock-market competition with scholarship prizes, and brought in speakers including a Wells Fargo VP.
FAI’s network: 6,000+ members · 126+ chapters · 30+ states · 4 countries
I’m the first author of a benchmark for how language models reason from incomplete security logs. I built ambiguous log scenarios, staged the release of new evidence, and evaluated how models revised their conclusions. Submitted to ACM SIGCITE 2026.
My poems The Geography of Us and The Ghost In The Gallery earned Scholastic Silver Keys in 2025 and 2026. My work also appeared in the American High School Poets anthology, and How to Live Inside a Wire was selected for the NFSPS Student Prize Winners book.
Alongside restoring cars, I ran ten automotive-literacy workshops for middle-school girls, including my sister and her friends. I showed them what I was repairing and how I was doing it, sharing the work as it happened.
USA Computing Olympiad · Platinum since February 2026
USACO Platinum
I reached Platinum, the highest division of the USA Computing Olympiad. The division focuses on advanced, open-ended algorithmic problems that require both a correct solution and an efficient implementation.
✳
Highest divisionBronze → Silver → Gold → Platinum
USACO advances contestants through four divisions based on contest performance.
01
My progression
I reached Silver in December 2024. On my first Silver attempt in February 2025, I scored 1000/1000 and advanced to Gold. I reached Platinum in February 2026.
02
What Platinum represents
USACO’s divisions progress from basic programming to advanced algorithms and data structures. Platinum is the highest level, with sophisticated, open-ended problems. Reaching it means earning promotion through the preceding divisions.
03
The problems I solve
My contest work includes dynamic programming, graph algorithms, and data structures under time pressure. The challenge is turning an unseen problem into an algorithm that is correct and fast enough for the time and memory limits.
How the map uses vector similarity to suggest related entries.
How it works +
How the ranking works
Each entry has three descriptions: the question it asks, the method it uses, and the constraints it works within. A sentence model turns each description into a vector. The map compares those vectors separately, then chooses a small set of relevant, less repetitive neighbors.
01 / REPRESENT
Separate embeddings
Question, method, and constraint are embedded separately, so shared reasoning can surface across different subjects.
02 / COMPARE
Weighted cosine similarity
A weighted combination of cosine scores changes what counts as a useful connection. Weak matches are filtered out.
03 / SELECT
Diversity ranking
A diversity penalty reduces redundant neighbors. Up to two links appear; the algorithm can return fewer when the evidence is weak.
SIMILARITY SCORES
USACO →
USAPhO
question
0.257
method
0.400
constraint
0.230
Weighted similarity 0.308
ROOK
question
0.145
method
0.271
constraint
0.379
Weighted similarity 0.242
Plain whole-description cosine
jaxphys · USAPhO
The baseline embeds each entry as one description. Scores measure similarity, not probability.
The implementation
Embeddings are generated ahead of time. Ranking runs locally on a fixed collection; there are no inference requests while you browse. The layout is composed by hand so points stay in place. The algorithm determines the related links.
The algorithm combines established sentence embeddings and diversity ranking with a custom three-part representation and scoring rules. The weights are fixed for this collection. The source includes the thresholds, comparisons, and tests.
