Solutions, notebooks and write-ups from competitive data science, kept as they were worked.
Competitions · Achievements · Courses · Toolbox · Structure · Using this repository · License
This repository holds the work behind a Kaggle profile: the notebooks entered into competitions, the reasoning that produced them, and the certificates and badges earned along the way. Each competition keeps its own folder, its own notebook and its own write-up explaining what the problem was, what was tried and where the approach ran out.
The write-ups are the point. A leaderboard score says how a submission placed and nothing about why, so each folder records the decisions: which signal the model was actually reading, which constraint shaped the design, and which of the three or four plausible approaches was taken and why.
Note
This is an active repository. Competitions are added as they are entered, and a folder is written when the work is done rather than when it succeeds. Some entries record approaches that did not place well and say so.
Tip
On trusting local validation over the public leaderboard. A public score is computed on a fraction of the test set, so it rewards models that fit that fraction's noise. If a change hurts local cross-validation but lifts the public score, it is usually the leaderboard that is wrong. Keeping a strict local setup and believing it is what holds a submission steady when the private split is revealed.
Important
The working method behind this repository, in one document: Kaggle Competitive Data Science
A companion that sets down what the competition folders only imply: the shared utilities of the Kaggle Toolbox, the notebook and write-up standard every folder follows, and the habits that carry from one competition to the next. It is this repository's own method written up, and is distinct from The Kaggle Book below, which is Konrad Banachewicz and Luca Massaron's.
The Kaggle Book · Konrad Banachewicz and Luca Massaron
Every folder carries its notebook and a write-up of the approach. The index is generated from the folders themselves, so it stays current as competitions are added.
| # | Competition | Medal | Kaggle |
|---|---|---|---|
| 1 | AI Mathematical Olympiad A modular agentic inference framework using RAG-backed symbolic computation and LaTeX-aware text diagnostics. |
Open | |
| 2 | Are You A Robot? Identifying AI-Generated Discourse through Stochastic Analysis |
Open | |
| 3 | BirdCLEF+ 2026 A high-performance bioacoustic pipeline using Perch v2 architecture and Bayesian model fusion for soundscape inference. |
Open | |
| 4 | CS Week Codeathon AIML (Easy Level) A predictive modeling approach using Exploratory Data Analysis (EDA) and Feature Engineering (FE) to forecast academic performance. |
Open | |
| 5 | Connect X A Minimax game-playing agent using Alpha-Beta pruning and heuristic board evaluation for the Connect Four simulation. |
🥉 | Open |
| 6 | English Scoring - Corrected Ver A predictive modeling approach using structured text features and ensemble regression to forecast student performance across seven scoring dimensions. |
Open | |
| 7 | Evading AI-Generated Text Detection Activation Steering and Lexical Variance for Detection Evasion |
Open | |
| 8 | AI Hallucination Visualizer A diagnostic pipeline for quantifying and visualizing stochastic uncertainty in LLMs using GPT-2 token-level probability analysis. |
Open | |
| 9 | Harmonizing the Data of your Data A predictive modeling approach using high-precision rule-based extraction and ontology normalization to structure scientific proteomics metadata. |
Open | |
| 10 | Hedge Fund - Time Series Forecasting Optimizing high-frequency investment signals through gradient boosted ensembles and multi-horizon temporal validation. |
Open | |
| 11 | House Prices A state-of-the-art regression pipeline using multi-model stacking, domain-driven feature science, and RMSLE-optimized ensembling. |
Open | |
| 12 | Kaggriculture An industrial livestock and predictive market-order prioritization engine in a two-player simulation. |
🥉 | Open |
| 13 | LLM Classification Finetuning Ensembled Pipeline Inference for Human Preference Classification |
Open | |
| 14 | Attention Span Evaluating Selective Attention and Distractor Vulnerability in Frontier LLMs. |
Open | |
| 15 | Petals to the Metal - Flower Classification on TPU Macro F1 Maximization through Distributed Dual-Stream Architectures |
Open | |
| 16 | Predict Customer Churn A gradient boosting ensemble pipeline combining Optuna-tuned XGBoost, LightGBM, and CatBoost with extensive feature engineering. |
🥉 | Open |
| 17 | Predicting Smartphone Addiction A Multi-Stage Meta-Feature Stacking architecture combining gradient boosted OOF interactions with Level-1 meta-learning for binary behavioural classification. |
Open | |
| 18 | Stanford RNA 3D Folding Part 2 Structural Biology Pipeline Optimization |
Open | |
| 19 | Student Study Hours to CGPA Prediction Predict academic performance using polynomial regression and regularized ensembles. |
Open | |
| 20 | Titanic - Machine Learning from Disaster A deterministic historical record linkage approach aligning passenger manifests with ground-truth survival records. |
🥉 | Open |
| 21 | Triagegeist A reliable three-tier clinical decision support system using a blended meta-ensemble and uncertainty-aware safety logic for ESI triage. |
Open |
Kaggle ranks each kind of contribution on its own scale, and this standing is in the Notebooks category: medals there are awarded by the community to published notebooks rather than for placing in a competition. Five of the notebooks kept here carry one: four competition notebooks, marked in the index above, and the shared Kaggle Toolbox.
| Badges | Medals | Tiers |
|---|---|---|
| Every badge earned, with its certificate | 9 bronze, awarded to published notebooks | Progression through the Kaggle ranks |
|
 |
 |
Note
Only the highest rank reached is recorded here. A live position moves every time anyone publishes, so a number written into a README is out of date the day after it is written. The Kaggle profile is the authority on where the standing sits today.
Seventeen Kaggle Learn courses completed, from introductory programming through to computer vision, time series and machine learning explainability. Every certificate is stored with the course it belongs to.
 1. Intro to Programming |
 2. Python |
 3. Intro to Machine Learning |
 4. Pandas |
 5. Intermediate ML |
 6. Data Visualization |
 7. Feature Engineering |
 8. Intro to SQL |
 9. Advanced SQL |
 10. Intro to Deep Learning |
 11. Computer Vision |
 12. Time Series |
 13. Data Cleaning |
 14. Intro to AI Ethics |
 15. Geospatial Analysis |
 16. ML Explainability |
 17. Game AI & Reinforcement Learning |
The helper module the notebooks share, so the same utilities are not pasted into the top of every one of them. Each function is there because something went wrong without it.
Tip
Zero-overhead integration in Kaggle Notebooks. Attach this module directly
under Notebook Settings → Add Utility Script or import via import kaggle_toolbox as tb.
It executes in sub-second time with zero external dependencies beyond standard scientific Python.
| Area | What it covers |
|---|---|
| Reproducibility | Locks every randomness source, including deterministic CUDA behaviour |
| Memory | Downcasts each column to the smallest safe dtype before a kernel runs out of RAM |
| Data quality | Missing-value reports, constant and duplicate column detection, correlation filtering |
| Submission safety | Checks row count, column order, duplicate IDs and prediction range before an attempt is spent |
| Diagnostics | Block timing, cross-validation summaries, hardware reporting and input path discovery |
Important
The failures that cost a competition never raise an exception: an unseeded run, a dtype that quietly exhausts memory, a file rejected after it has already spent an attempt.
Seed before you compare. Validate before you submit. Each function here is one of those checks, cheap enough to run every time.
– Amey Thakur
It ships with a demo notebook and a tutorial script. Full reference →
.
├── Competitions/ # One folder per competition, with its notebook and write-up
├── Achievements/
│ ├── Badges/ # Every badge earned, with its certificate
│ ├── Medals/ # Medal artwork
│ └── Tiers/ # Progression tier artwork
├── Kaggle Courses/ # Course certificates
├── Kaggle Toolbox/ # Shared helper module used across the notebooks
├── docs/ # Repository imagery
├── CITATION.cff # How to cite this work
├── codemeta.json # Machine-readable project metadata
├── LICENSE # CC BY 4.0, for written material
└── LICENSE-MIT # MIT, for code
Each competition folder is self-contained. Open its README.md for the write-up
and the notebook alongside it for the implementation, or use the Open in
Kaggle badge to run it there with the data already attached.
Important
Most notebooks depend on Kaggle datasets attached in the notebook environment and on competition data that cannot be redistributed here. They are written to run on Kaggle. Cloning the repository gives you the code and the reasoning, not a runnable environment.
Written material, including every README and write-up, is released under Creative Commons Attribution 4.0 International. Code, including the notebooks and the toolbox, is released under the MIT License.
Copyright © 2026 Amey Thakur