I am a sophomore in Fremont, and I spend most of my time in other people's codebases. Not writing tutorials about them. Fixing them. When I'm not doing that, I design robots for my FTC team and build AI agents at hackathons.
- 132 merged pull requests into 50 repositories across 25 organizations, including Meta, Apple, NVIDIA, OpenAI, Google and the Rust project.
- 4 hackathon wins in 2026.
- Lead CAD Designer and Design Captain of FTC Team 23543, The Architects. 3x FTC Qualifier Finalist.
- Qualified for the AIME through the AMC 10.
| Hackathon | Result | Build |
|---|---|---|
| Beta Fund x EverMind Super Hackathon May 2026 |
1st place and Audience Favorite, $600 | SwarmMarket, an app store for agent experience. One weekend, one teammate. |
| Beta Fund AI Super Hackathon May 2026 |
Winner, $500 | ShadowBuyer, a six-agent procurement system that negotiates with vendors. |
| Eco Hacks April 2026 |
Winner | EcoLens, which looks at a photo of old electronics, estimates what they're worth, and tells you whether to sell, trade in or recycle. |
| IHS CSDECA Hackathon November 2026 |
Winner, team of two | ShadowBuyer for Main Street, AI agents that get restaurant owners their card-processing money back, over text. |
|
A restaurant owner texts a photo of their card-processing statement. Claude agents find the hidden fees, expose a "1.35 factor" cash advance as about 93% APR, and negotiate with a simulated processor rep. Nothing happens until the owner replies YES, and every dollar figure comes from unit-tested code, never the model. |
Agents publish what they learned as portable skill packs, and other agents install those lessons in one call, with no retraining. A React and Vite marketplace shows live before-and-after behavior, on a Node backend built on the Claude API with an MCP server. |
|
Semantic search over 1,319 winning Devpost projects from 408 hackathons, packaged as a Claude Code skill. Describe the hackathon you're entering and it pulls the closest past winners, then turns them into patterns and ideas. ChromaDB, sentence embeddings, 26 pytest tests. |
A calibrated expected field goal percentage model on free playoff data. The point is the method: cross-validation grouped by game so nothing leaks between folds, target encoding inside the pipeline, a logistic-regression baseline the bigger model has to beat, and empirical-Bayes shrinkage so a player with 25 shots can't top the rankings. |
|
Thirty NBA analytics tools behind one natural-language command bar: shot quality, player similarity, a cap-legal trade machine, lineup optimization. Every panel computes live from ESPN data for 500+ rostered players. 158 engine unit tests and a hand-built SVG chart kit with zero charting dependencies. |
Six agents handle procurement from start to finish: vendor scouting, quote collection, adversarial negotiation, contract redlines and email drafting. In the demo negotiation it cut the price from $195 to $157.50 per host per month. Python, FastAPI and React. |
132 merges. 25 organizations. 50 repositories. I sent every one cold: no referral, no internship, nobody who knew my name first.
The pitch is narrow on purpose. I find a real bug in a library you already depend on, write the regression test that fails before and passes after, and send it in.
|
pyrefly#3415 fixes negative narrowing for |
NVIDIA's maintainer opened cuda-python#2410 to move the project from |
The ten I would defend in a code review (click to open)
| Org | Pull request | What it actually fixes |
|---|---|---|
| Meta | pyrefly#3415 | Negative narrowing for isinstance tuple targets. Landed through Phabricator; mypy_primer classified it as 5 improvements and 4 neutral across 9 projects. |
| websockets | ws#2329 | A maxFragments limit bypass: empty continuation frames slipped past the counter. |
| Rust | rust#156787 | A comment correction in alloc, on the ArcInner::weak lock sentinel. Small, but it's in the standard library. |
| Meta | lexical#8995 | Breaking change to the clipboard payload: excludeFromCopy was being asked about the wrong destination. One of 36 merges into Lexical. |
| Meta | lexical#8918 | Deprecated CAN_UNDO_COMMAND and CAN_REDO_COMMAND in favor of HistoryExtension signals. An API direction change, not a patch. |
| Apple | coremltools#2700 | K-means weight palettization was scoring centroid sums instead of intra-cluster error, so quantized models came out worse than they had to. |
| Apple | password-manager-resources#1094 | A CI check that parses every password rule, so a malformed rule fails before it reaches browsers and password managers. |
| NVIDIA | TransformerEngine#3065 | skip_fp8_weight_update was not reaching GroupedLinear during FP8 CUDA graph capture. Added a CUDA graph test. |
| Vue | core#14865 | Persisted transition hooks fired on every KeepAlive cache hit. Fixed in runtime-core, with a regression test. |
| OpenAI | openai-agents-python#3485 | Raw failed JSON tool input was echoed into ModelBehaviorError even with tool-data logging off, a quiet way to leak whatever the model was handed. |
The full docket, by house (25 organizations)
| House | Repositories | Merges |
|---|---|---|
| Meta | lexical · pyrefly · jscodeshift · stylex · idb |
42 |
| Apple | coremltools · container · containerization · password-manager-resources · swift-nio · servicetalk · foundationdb |
20 |
| NVIDIA | cuda-python · TransformerEngine · earth2studio · cutlass · apex · Megatron-LM · gpu-operator |
20 |
| OpenAI | openai-agents-python · openai-agents-js · openai-node · openai-python |
10 |
| Microsoft | rushstack · rushstack-websites · PowerApps-Samples |
6 |
go-cloud |
5 | |
| Fastify | fastify · fastify-static · fastify-autoload · light-my-request |
4 |
| Deno | deno · std |
4 |
| Cloudflare | workers-sdk |
4 |
| Jest | jest |
2 |
| Others | rust-lang/rust · vuejs/core · vitejs/vite · rollup/rollup · websockets/ws · vercel/satori · honojs/hono · h3js/h3 · Kludex/starlette · nanostores/nanostores · ml-explore/mlx · huggingface/datasets · unjs/unstorage · open-circle/valibot · nestjs/nest |
15 |
Counted from merged pull requests authored by LeSingh1, excluding my own repositories. Two Meta contributions (pyrefly and idb) landed through Phabricator, which GitHub does not label as merged, so they are added by hand.
How I actually find these
No fuzzing, no scanners, no bulk codemods.
- Read the issue tracker backwards. Old open issues with a clear repro and no assignee are the best inventory in open source. Somebody already did the hard part of proving a bug exists.
- Check that two code paths agree. Most of my bugs come from asking whether two paths that should behave the same actually do: streamed and non-streamed, sync and async, a base class and its override, Linux and Windows.
- Read the tests, not the code. The gap between what a test file covers and what the module claims to do is where bugs sit.
toMatchObjectnot handling class instances was sitting in that gap. - Write the failing test first. If I cannot make it go red, I do not understand the bug well enough to send a patch.
- One behavior per PR. Maintainers are volunteers. A reviewable diff is a merged diff.
I'm the Lead CAD Designer and Design Captain for FTC Team 23543, The Architects, a 3x FTC Qualifier Finalist. I design our chassis, mechanisms, layout and blueprints in Fusion 360.
A digital twin of our robot. This fall I built a model of our competition robot in Fusion 360: about 375 parts, down to every fastener, modeled from about 30 photos and driven by Python scripts through the Fusion API. I exported it into a MuJoCo physics simulation and ran intake, transfer, shooting and drive tests. In simulation, our intake could not pick a ball up off the floor. From those findings I directed a redesign with funnel guides that center each ball, and in simulation it passed every ball case: 59 of 59 starting positions and 64 of 64 ball sizes, inside the 18-inch size limit.
Next: a sense of touch. I'm planning a research project on one question: can a cheap contact microphone hear a robot gripper start to slip, earlier and more reliably than a force sensor? The plan is an FTC gripper with a contact mic and a force sensor side by side, measuring how many milliseconds of warning each one gives before a drop.
| Basketball | Irvington High School. Captain of the freshman team in 2025-26, where I led the team in scoring, rebounding and blocks. Now on varsity as the only sophomore. |
| Senior Patrol Leader | Scouts, Troop 199. I lead a troop of 40 to 50 scouts: weekly meetings, outings and service projects. Life Scout. Service work I've done includes restoring a pony pen at Ardenwood Historic Farm, building a retaining wall along the trail at Tule Ponds, building a shed to protect the electrical equipment and water pump at the Warm Springs Little League snack shack, and helping build 20 patlas, the small tables used for poojas. |
| Lead CAD Designer | FTC Team 23543, The Architects. Fusion 360: chassis, mechanisms, layout and blueprints. 3x FTC Qualifier Finalist. |
| AIME Qualifier | Qualified for the AIME through the AMC 10. |
| Anthropic Academy | Courses on Claude Code, agent skills and subagents. May 2026. |
Fremont, California · Class of 2029
Artwork generated by art/build_art.py: a 1st Edition holo trading card on a graffiti wall, in Philadelphia blue and red with white and silver foil. HP is the merge count and every move carries a real number.
Every shape is original and every glyph is a real vector outline, because GitHub renders these inside <img> and an <img> cannot load a font.



