Skip to main content
IST --:--:--
ExperienceStackExtraEeshOS3D
LinkedIn ↗GitHub ↗Kaggle ↗Codolio ↗
(02) KaggleKaggle profile ↗

A silver and a bronze so far, from 280 submissions and hundreds of hours, with work across medical imaging, coding agents, agent security, 3D cell tracking and game-playing agents.

473/10,148

Kaggriculture Silver

Solo, 99 submissions, top 4.7%. Offline wins against recorded games did not survive a live, reacting ladder.

Write-up ↗
287/3,947

Biohub Tracking Bronze

Team of four, 181 submissions, top 7.3%. Sub-voxel soft-argmax localization was the biggest gain.

Write-up ↗
135/6,094

RSNA Knee ● Live

Medical imaging, live public leaderboard. Ends October 22.

Competition ↗
(01) How I approach a competition
  1. 01

    Read the metric

    Work out exactly what the score rewards and punishes before writing a model.

  2. 02

    Validation first

    Build a leak-free validation setup that predicts the leaderboard, not one that flatters me.

  3. 03

    Honest baseline

    Get a simple, honest baseline on the board early so every later idea has something to beat.

  4. 04

    One idea that moves it

    Spend the rest of the time on the one modelling idea that actually moves the metric.

(02) Worked on so far7 competitions
01

Kaggriculture

A two-player farming-simulation agent ladder, ranked by a Bradley-Terry fit over games played after the deadline. My agents build on the public Shepherd's Ledger tape-replay chassis with layers of our own, and the research was about which edits to a replay agent actually move win rate in a shared-market economy. The biggest lesson: every layer measured as a gain offline against recorded games, yet the later agents rated lower on the live ladder, because the gains were small next to the game-to-game swing and recorded opponents do not react. Solo entry, 99 submissions over four weeks.

Silver · 473 / 10,148

Game AI · Agents · Simulation

Write-up ↗
02

Biohub Cell Tracking During Development

3D nuclei tracking in developing zebrafish. Two pipelines share a detector and tracker but place coordinates differently, so they fail independently on the embryo-disjoint hidden set; sub-voxel soft-argmax localization was the single biggest gain. Built as a team of four over two months, with 181 submissions.

Bronze · 287 / 3,947

Computer Vision · 3D · Tracking

Write-up ↗
03

RSNA Knee Abnormality Detection

Detecting knee abnormalities in a kernels-only medical-imaging competition with 6,000+ teams. I blend fine-tuned vision models and check every change against a pre-registered evaluation before it ships.

Ongoing · ends Oct 22

Medical Imaging · PyTorch

04

Google - The Gemma 4 Developer Agent

Building a coding agent on Gemma 4 that resolves real software-engineering tasks in private repositories, within a 12-hour budget for the whole hidden test set.

Ongoing · ends Dec 2

Coding Agents · LLM · Gemma

05

AI Agent Security: Multi-Step Tool Attacks

The attack track: crafting multi-step prompt injections that make a tool-using LLM agent exfiltrate data past its guardrails. My best public score was 106.285, and I published a working note walking through the approach.

Completed

LLM Security · Prompt Injection · Agents

Working note ↗
06

Pokemon TCG AI Battle

A bot-ladder competition for the Pokemon Trading Card Game. I built a Monte Carlo Tree Search agent tuned around the scoring quirks of the ladder and its search-depth trade-offs.

Completed

Game AI · MCTS

07

ARC Prize 2026 (ARC-AGI-2)

The abstract-reasoning benchmark. I explored test-time training and program synthesis on small open models under the offline compute budget, then paused it to focus on the RSNA competition.

Paused

Reasoning · Test-Time Training · LLM