Singapore-based practical guides, tutorials and experiments in AI, computing, modelling, simulation, optimisation and quantum computing, with research notes and hands-on workflows.
POCKET DATA SCIENCE SERIES 2026 The complete mobile machine learning roadmap: from Termux PRoot bootstrapping to Global Top 100 on Kaggle.
This is the master index and curriculum for the Pocket Data Science series: a sequence of hands-on guides and empirical benchmarks proving that competitive machine learning, feature engineering, and statistical modeling can be executed entirely on an Android mobile device. Using Android Termux, isolated PRoot Linux containers, and the Google Antigravity CLI (agy), team Pocket Data Lab designs, tests, and deploys high-performing models directly from handheld edge hardware. Each article stands on its own, but together they form a systematic arc—from mathematical foundations and mobile environment bootstrapping, to tabular classification, target geometry, mobile vision and NLP, time-series quad-ensembles, and LLM human preference alignment. The order below is a learning path from first principles to advanced competition rank, rather than the chronological order of publication.
New to edge data science? Read top to bottom starting from Foundational Pre-reading. Already comfortable with loss math and mobile PRoot setup? Jump straight to tabular classification in Track 2 or the quad-ensembles in Track 5. Explore all competition code on the official Pocket Data Lab Kaggle Profile.
Foundational Pre-reading: The Math Behind Loss & Evaluation Metrics
The theoretical foundation of probability calibration: deriving cross-entropy from Maximum Likelihood Estimation, why confident false predictions cause catastrophic infinite penalties, probability calibration, and numerical clipping bounds (eps = 1e-15).
Evaluation metric taxonomy: understanding type I and type II error trade-offs under class imbalance, macro vs. micro vs. weighted F1 formulations, and probability threshold tuning on mobile edge models.
Step-by-step installation of Ubuntu/Debian inside Termux PRoot on Android ARM64: building native wheels for numpy, scipy, and pandas, configuring the Google Antigravity autonomous coding agent, setting up Git submodules, and establishing headless Kaggle API token authentication.
Edge documentation pipelines: automating dispatch and research reporting directly from mobile terminal shells without opening browser administration panels.
Foundational tabular classification: title parsing, demographic imputations, out-of-fold blended ensembling (Random Forest, LightGBM, CatBoost), and automated Kaggle dispatch securing an official Top 3% global finish (Score: 0.79665).
Exploiting problem-domain physics: CryoSleep spending invariants, passenger group surname matching, and native CatBoost on ARM64 achieving an official Top 6% global standing (Score: 0.80687).
Computer vision under tight memory constraints: PCA orthogonal subspace projection, Support Vector Machines (SVM) with radial basis kernels, and training 42,000 samples within a 2 GB Android RAM budget.
Natural language processing on edge devices: n-gram TF-IDF sparse matrices, text tokenization, and regularized logistic baselines proving competitive against heavy transformer architectures.
Deterministic time-series forecasting: building hierarchical rolling averages and holiday heuristics without machine learning weights, outperforming 66% of competitors (RMSLE: 0.44975).
The capstone forecasting architecture: combining moving averages, calendar regression, and GBDT residual models into a convex 4-way blend, clinching Global Rank #99 on Kaggle (RMSLE: 0.40455).
6 · Generative AI Alignment & LMSYS Human Preference
Predicting human preference across pairs of generative LLM outputs: exploiting length bias heuristics, probability distribution tuning for multi-class log loss, and placing Rank #222 out of 291 teams with zero neural weights.
All models, feature engineering pipelines, and cross-validation folds below were trained and dispatched entirely from an Android device via Termux PRoot Linux:
Competition & Module
Methodology & Stack
Kaggle Standing
Titanic (Part I)
10-Fold GBDT Blended Ensemble
Top 3% Global (0.79665)
Spaceship Titanic (Part II)
CryoSleep Invariants & Native CatBoost
Top 6% Global (0.80687)
House Prices (Part III)
Log-Transformed Regularized Regression
Benchmarked (RMSLE)
MNIST Digits (Part IV)
PCA Subspace Projection + Support Vector Machine
RAM Bound (2 GB Budget)
Disaster Tweets (Part V)
Sparse TF-IDF N-grams + Linear Models
Edge NLP Baseline
Store Sales (Part VI)
Hierarchical Non-ML Moving Averages
Top 33.1% Global (0.44975)
Store Sales (Part VII)
Deterministic Trend + LightGBM Residuals
Top 21.0% Global (0.41843)
Store Sales (Part VIII)
Convex 4-Way Quad-Ensemble Blend
Global Top 100 · Rank #99 (0.40455)
LMSYS Baselines (Part IX)
Calibrated Response Length Difference Heuristic
Rank #222 of 291 (Log Loss 1.07823)
LMSYS GBDT ML (Part X)
58 Linguistic Feats + Dual GBDT Ensemble (LGB + CB)
Rank #170 of 290 (Log Loss 1.04158)
Pocket Data Science Series 2026 · Engineered on Android Termux & PRoot Linux
✦ Benchmarked and verified by team Pocket Data Lab on Kaggle ✦
Leave a Reply